NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 1 OF 52
NISTIR 7779
An Exploration of the Operational
Ramifications of Lossless Compression of
1000 ppi Fingerprint Imagery
Shahram Orandi
John M. Libert
John D. Grantham
Kenneth Ko
Stephen S. Wood
Jin Chu Wu
Lindsay M. Petersen
Bruce Bandini
http://dx.doi.org/10.6028/NIST.IR.7779
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 2 OF 52
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 3 OF 52
NISTIR 7779
An Exploration of the Operational
Ramifications of Lossless Compression of
1000 ppi Fingerprint Imagery
Shahram Orandi
John M. Libert
Kenneth Ko
Stephen S. Wood
Jin Chu Wu
Information Access Division - Image Group
Information Technology Laboratories
John D. Grantham
Systems Plus, Inc.
Lindsay M. Petersen
MITRE Corporation
Bruce Bandini
Booz Allen Hamilton, Inc.
http://dx.doi.org/10.6028/NIST.IR.7779
August 2012
U.S. Department of Commerce
Rebecca Blank, Acting Secretary
National Institute of Standards and Technology
Patrick D. Gallagher, Under Secretary of Commerce for Standards and Technology and Director
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 4 OF 52
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 5 OF 52
ACKNOWLEDGEMENTS
The authors wish to give special thanks to the following individuals and organizations for their support of this work:
Federal Bureau of Investigation for all their support throughout this study
T.J. Smith and the LA County Sheriff’s Department
R. Michael McCabe, IDTP
Margaret Lepley, MITRE
In addition, we are grateful for and appreciate the guidance, support and coordination provided by Michael D. Garris,
without whose help this study would not have been possible.
DISCLAIMER
Specific hardware and software products identified in this report were used in order to perform the evaluations
described in this document. In no case does identification of any commercial product, trade name, or vendor, imply
recommendation or endorsement by the National Institute of Standards and Technology, nor does it imply that the
products and equipment identified are necessarily the best available for the purpose.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 6 OF 52
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 7 OF 52
EXECUTIVE SUMMARY
The criminal justice communities throughout the world exchange fingerprint imagery data primarily in 8-bit gray-scale
and at 500 pixels per inch1
(ppi) or equivalently 19.7 pixels per millimeter (ppmm). The Wavelet Scalar Quantization
(WSQ) fingerprint image compression algorithm has been developed and maintained by the Federal Bureau of
Investigation (FBI), Los Alamos National Laboratory and the National Institute for Standards and Technology (NIST) as
the standard for compressing 500 ppi fingerprint imagery in the United States. WSQ is classified as a “lossy”
compression algorithm. Lossy compression algorithms employ data encoding methods which discard (lose) some of the
data in the encoding process in order to achieve an aggressive reduction in the size of the data being compressed.
Decompressing the resulting compressed data yields content that, while different from the original, is similar enough to
the original that it remains useful for the intended purpose. The WSQ algorithm allows users to specify how much
compression is to be applied to the fingerprint image at the cost of increasingly greater loss in fingerprint image fidelity
as the effective compression ratio is increased (see Figure 1 for an example of image degradation from lossy
compression).
The importance of latent images (images lifted from the crime scene, via uncontrolled collection) in criminal casework
has been growing. Anecdotal evidence from latent fingerprint examiners has indicated that any fidelity loss as a result
of compression should be avoided prior to review of those fingerprints by examiners and that even small increases in
the amount of higher level detail may yield benefits. Because of this, most latent fingerprint images used in casework
are either transferred in non-compressed form or compressed using a lossless algorithm where the resulting
compressed representation of the image can be decompressed to yield an image exactly identical to the original.
While extensive effort has been put into standards and certification pathways for lossy fingerprint compression,
lossless compression strategies have not been thoroughly examined.
This study examines various lossless compression algorithms and the relative advantages and disadvantages of each
with respect to effective compression rates, compression throughput and decompression throughput. This study also
examines different implementations of the same algorithm, as well as implementations of the same algorithm
generated for different computer system architectures (e.g., 32-bit & 64-bit).
This study finds that wavelet-based compression algorithms such as JPEG 20002
generally yield better effective
compression rates than the non-wavelet-based algorithms such as PNG (Portable Network Graphics). However, nonwavelet-based algorithms tend to have higher throughput (require less time to operate on data) than wavelet-based
algorithms. This study also shows that while architectural differences do not yield much operational difference in terms
of effective compression rates, such architectural differences do translate to significant differences in compression and
decompression throughputs.
1Resolution values for fingerprint imagery are specified in pixels per inch (ppi) throughout this document. This is based on widely
used specification guidelines for such imagery and is accepted as common nomenclature within the industry. SI units for these will be
presented only once.
2
The “2000” refers to the year of publication of the first edition of the image compression standard known as JPEG 2000. JPEG
refers to Joint Photographic Experts Group.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 8 OF 52
VERSION HISTORY
Date Activity
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 9 OF 52
TABLE OF CONTENTS
1. Investigative Goals and Objectives .......................................................................................................................................................17
1.1. Background ...................................................................................................................................................................................17
1.2. Key Drivers and Mandates ............................................................................................................................................................18
2. Materials and Methods..........................................................................................................................................................................19
2.1. Compression Algorithms ..............................................................................................................................................................19
2.2. Compression Conditions ...............................................................................................................................................................21
3. Analysis.................................................................................................................................................................................................. 22
3.1. Performance Metrics ................................................................................................................................................................... 22
3.2. Uncertainty Measurement .......................................................................................................................................................... 23
3.3. Algorithmic Comparisons ............................................................................................................................................................ 24
3.4. Architecture and Implementation Comparisons ........................................................................................................................ 28
4. Results....................................................................................................................................................................................................31
4.1. Investigative Goal 1: Examine Effective Compression Rates by Algorithm ................................................................................31
4.2. Investigative Goal 2: Examine Compression Throughput by Algorithm .................................................................................... 32
4.3. Investigative Goal 3: Examine Decompression Throughput by Algorithm................................................................................ 33
4.4. Investigative Goal 4: Examine Effective Compression Rate by Impression Type...................................................................... 34
4.5. Investigative Goal 5: Comparison of JPEG 2000 Implementation With Respect to Effective Compression Rate ................... 35
4.6. Investigative Goal 6: Comparison of JPEG 2000 Implementation With Respect to Compression Throughput ......................36
4.7. Investigative Goal 7: Comparison of JPEG 2000 Implementation With Respect to Decompression Throughput................... 37
4.8. Investigative Goal 8: Comparison of Implementation Complexity ............................................................................................38
5. Conclusions ...........................................................................................................................................................................................39
6. Discussions and Future Work ...............................................................................................................................................................40
References ......................................................................................................................................................................................................41
Publications and Reports ...........................................................................................................................................................................41
Standards................................................................................................................................................................................................... 43
Appendix A. Dataset Makeup ..................................................................................................................................................................45
Appendix B. Equipment Used for Study..................................................................................................................................................50
Appendix C. Examination of Entropy .......................................................................................................................................................51
Procedure ...................................................................................................................................................................................................51
Results ....................................................................................................................................................................................................... 52
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 10 OF 52
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 11 OF 52
LIST OF TABLES
Table 1 - Abbreviations.............................................................................................................................................................. 15
Table 2 –Impression Types Selected for This Study ................................................................................................................19
Table 3 - Compression Conditions.............................................................................................................................................21
Table 4 - Effective Compression Rates (Medians), Higher is Better....................................................................................... 31
Table 5 – Effective Compression, Wilcoxon Signed Rank Test, alpha of 0.05 ....................................................................... 31
Table 6 - Compression Throughput (Median Time, seconds), Lower is Better .....................................................................32
Table 7 - Compression Throughput, Wilcoxon Signed Rank Test, alpha of 0.05 ...................................................................32
Table 8 - Decompression Throughput (Median Time, seconds).............................................................................................33
Table 9 - Decompression Throughput, Wilcoxon Signed Rank Test, alpha of 0.05...............................................................33
Table 10 - Algorithm Performance Rankings (1-Best through 6-Worst).................................................................................34
Table 11 - Effective Compression Rates (Medians) by Implementation, Higher is Better.....................................................35
Table 12 - Effective Compression by Implementation, Wilcoxon Signed Rank Test, at alpha = 0.05....................................35
Table 13 - Compression Throughput (Median Time, seconds) by Implementation, Lower is Better ...................................36
Table 14 - Compression Throughput by Implementation, Wilcoxon Signed Rank Test, at alpha = 0.0033 ..........................36
Table 15 - Decompression Throughput (Median Time, seconds) by Implementation, Lower is Better ...............................37
Table 16 - Decompression Throughput, Wilcoxon Signed Rank Test, alpha of 0.05 .............................................................37
Table 17 - Implementation Complexity ....................................................................................................................................38
Table 18 - Ink Card Scan Data classification by Impression Type ............................................................................................45
Table 19 - Ink Card Scan Pattern Classification for Single Finger Images by Impression Type..............................................45
Table 20 - Ink Card Scan Pattern Classification for Single Finger Images by Finger (Females)............................................ 46
Table 21 - Ink Card Scan Pattern Classification for Single Finger Images by Finger (Males) ................................................ 46
Table 22 – Digital Live Scan Data Classification by Impression Type ..................................................................................... 46
Table 23 – Digital Live Scan Pattern Classification for Single Finger Images by Impression Type........................................47
Table 24 – Digital Live Scan Pattern Classification for Single Finger Images by Finger (Females) .......................................47
Table 25 – Digital Live Scan Pattern Classification for Single Finger Images by Finger (Males) ...........................................47
Table 26 - Gender Breakdown for Data .................................................................................................................................. 48
Table 27 - Age Breakdown for Data ........................................................................................................................................ 48
Table 28 - Other Metadata: Height and Weight ..................................................................................................................... 48
Table 29 - Other Metadata: Eye Color..................................................................................................................................... 48
Table 30 - Image Geometry Data............................................................................................................................................. 49
Table 31 – Mean measurements of entropy and fingerprint region of images .....................................................................52
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 12 OF 52
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 13 OF 52
LIST OF FIGURES
Figure 1 - Example of Fidelity Degradation Due to Extreme Lossy Compression (JPEG 2000 at 800:1)...............................18
Figure 2 - Normal Fit Histogram, Effective Compression (PNG)............................................................................................. 23
Figure 3 - Q-Q Normality Plot, Effective Compression (PNG) ................................................................................................. 23
Figure 4 - Effective Compression Rates (Medians)................................................................................................................. 25
Figure 5 - Compression Throughput (Median Time, seconds)................................................................................................26
Figure 6 - Decompression Throughput (Median Time, seconds............................................................................................. 27
Figure 7 - Effective Compression Rate (Median Ratio, by Platform and Implementation)...................................................28
Figure 8 - Compression Throughput (Median Time, seconds, by Platform and Implementation).......................................29
Figure 9 - Decompression Throughput (Median Time, seconds, by Platform and Implementation)...................................30
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 14 OF 52
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 15 OF 52
TERMS AND DEFINITIONS
The abbreviations and acronyms of Table 1 are used in many parts of this document.
Table 1 - Abbreviations
BMP Bitmap File Format
bpp Bits per pixel
CR Compression Ratio
FBI Federal Bureau of Investigation
IAFIS Integrated Automated Fingerprint Identification System
IAI International Association for Identification
ICER Image Compression and Stereo Ranging
JPEG Joint Photographic Experts Group – ISO/IEC committee developing standards for image
compression – also used as the name of the CODEC developed in accordance with the
standard specified by this body.
NBIS NIST Biometric Image Software
NGI Next Generation Identification
NIST National Institute of Standards and Technology
OPJ OpenJPEG’s JPEG 2000 CODEC
PNG Portable Network Graphics
ppi Pixels per inch
ppmm Pixels per millimeter
PSNR Peak Signal To Noise Ratio
RLE Run Length Encoding
SIVV Spectral Image Validation/Verification Metric
WSQ Wavelet Scalar Quantization algorithm for compression of fingerprint imagery
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 16 OF 52
ABSTRACT
This paper presents the findings of a study initially conducted to measure the operational impact of JPEG 2000 lossy
compression on 1000 ppi fingerprint imagery at various levels of compression, but later expanded to include lossless
compression. Lossless compression will have no impact on either Galton or non-Galton based features of a fingerprint
since the compressed image is identical to the original once decompressed. The selection of a lossless compression
algorithm can have operational implications in terms of effective compression rate and throughput; these implications
are the focus. This study examines several such compression algorithms and compares them using criteria used to
measure the effectiveness of the compression algorithm as well as its throughput using actual fingerprint imagery.
KEYWORDS
Fingerprint compression; 1000 ppi fingerprint imagery; JPEG 2000; JasPer; OpenJPEG; PNG; RLE; BMP; ICER; lossless
compression
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 17 OF 52
1. Investigative Goals and Objectives
In July of 2009 NIST in partnership with the FBI commenced an investigation on the use of JPEG 2000 [JPEG2K] for
compressing fingerprint imagery. Part one of this investigation, described in [NISTIR7778], addressed JPEG 2000 when
operating with a non-reversible/lossy filter. The study described in this paper complements [NISTIR 7778] by examining
the performance of several compression algorithms including JPEG 2000 when operating in a lossless fashion with the
following investigative goals:
1. Examine Effective Compression Rates by Algorithm: Assess the performance of the selected algorithms
according to how much effective compression they yield for various impression types.
2. Examine Compression Throughput by Algorithm: Assess the performance of the selected algorithms with
respect to the time needed to generate a compressed representation of the original image.
3. Examine Decompression Throughput by Algorithm: Assess the performance of the selected algorithms with
respect to the time needed to reconstruct the original image from the compressed representation.
4. Examine Effective Compression Rate by Impression Type: Identify if a particular impression type impacts
effective compression rates more than other types.
5. Compare JPEG 2000 Implementations with Respect to Effective Compression Rate: Identify any operational
differences in effective compression rate between various implementations of the JPEG 2000 algorithm for the
same given set of input images.
6. Compare JPEG 2000 Implementations with Respect to Compression Throughput: Identify any operational
differences in compression throughput between various implementations of the JPEG 2000 algorithm for the
same given set of input images.
7. Compare JPEG 2000 Implementations with Respect to Decompression Throughput: Identify any operational
differences in decompression throughput between various implementations of the JPEG 2000 algorithm3
for
the same given set of input images.
8. Compare Implementation Complexity: Examine and compare algorithm codebase size and complexity.
As an ancillary component of this study, the compression/decompression pathway was empirically verified to be truly
lossless where the resulting compressed image stream would yield the exact original image, pixel-for-pixel.
1.1. Background
The criminal justice communities throughout the world exchange fingerprint imagery data primarily in 8-bit gray-scale
and at 500 pixels per inch (ppi). The Wavelet Scalar Quantization (WSQ) [BRADLEY1], [BRADLEY2], [BRISLAWN],
[HOPPER] image compression algorithm has been developed and maintained by the Federal Bureau of Investigation
(FBI), Los Alamos National Laboratory (LANL) and the National Institute for Standards and Technology (NIST) as the
standard for compressing 500 ppi fingerprint imagery in the United States. The WSQ standard defines a class of
encoders and decoders with sufficient interoperability to ensure that images encoded by one compliant encoder can be
decoded by any other compliant decoder.
WSQ is a “lossy” compression algorithm. Lossy compression algorithms employ data encoding methods that discard
(lose) some of the data in the encoding process in order to achieve an aggressive reduction in the size of the data being
compressed. Decompressing the resulting compressed data yields content that, while different from the original, is
similar enough to the original that it remains useful for the intended purpose. Lossless compression algorithms, on the
other hand, produce a compressed image that can be decompressed back to original form with no loss or change to the
image. The disadvantage to lossless algorithms is that they produce compressed images that can be many times larger
in file size than compressed images produced by lossy algorithms.
3
For purposes of the present study, both compression and decompression were performed by the same JPEG 2000 implementation.
That is, there was no cross-over among implementations, e.g. encoding with one and decoding with another CODEC.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 18 OF 52
The WSQ algorithm allows users to specify how much compression to apply, where higher amounts of compression
result in a more compact representation but greater loss in fidelity to the original image.
Figure 1 below shows an example of such image degradation and fidelity loss as a result of lossy compression. The WSQ
Gray-Scale Fingerprint Image Compression Specification [WSQ] provides guidance for the acceptable amount of fidelity
loss due to compression in order for the encoder and decoder to meet FBI certifications for 500 ppi fingerprint imagery.
These certifications are designed to ensure adherence to the WSQ specification to ensure sufficient fidelity for
admissibility of images in courts of law that have been processed by such encoders and decoders.
Original Image Compressed Image (Lossy)
Figure 1 - Example of Fidelity Degradation Due to Extreme Lossy Compression (JPEG 2000 at 800:1)
A study conducted by the International Association for Identification (IAI) [FITZPATRICK] established 15:1 as a WSQ
compression ratio that would retain acceptable image fidelity in 500 ppi fingerprint imagery. The study used the
judgments of expert fingerprint examiners to measure the fidelity loss due to compression. In order to reduce bias due
to subjectivity, multiple examiner decisions were used to build consensus. Utilizing examiners’ opinions does not imply
that automated fingerprint matcher performance is not an important criterion in a given biometric system, but it must
be noted that if fingerprints are to be admissible as evidence in a court of law their ultimate utility lies in the expert
examiner’s opinion of the fidelity of those fingerprints.
1.2. Key Drivers and Mandates
In modernizing its environment as part of the Next Generation Identification (NGI) initiative, the FBI seeks to expand its
ability to exchange fingerprints at 1000 ppi in an effort to improve upon the capacity of systems in fingerprint
identification and verification tasks and meet the FBI mandate to:
- Protect the United States from terrorist attack, foreign intelligence operations and espionage
- Support federal, state, local and international partners in their efforts to prevent or reduce crime and violence
- Upgrade technology to support the FBI's missions
Toward meeting these goals, the FBI seeks to set guidance for the next generation encoders and decoders based on
the JPEG 2000 compression standard [JP2K] in order to ensure interoperability, fidelity and admissibility for 1000 ppi
images in courts of law in the criminal justice community. While this is the case for lossy compression, other CODECs
might be considered for lossless compression of fingerprints for which the desire is to maintain absolute fidelity of the
image to the original, yet achieve some savings of storage space.
In support of the FBI, the National Institute of Standards and Technology (NIST) conducted a study to determine an
optimal compression approach that follows on the IAI study of WSQ compression for 500 ppi fingerprint imagery, to
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 19 OF 52
build upon existing guidance for JPEG 2000 compression of fingerprint imagery, and to formulate a basis with which a
normative compression guidance can be established in the ANSI/NIST standard for biometric data interchange
[AN2011].
NIST has an established expertise in evaluating biometric systems and standards, and has been assigned by the USA
PATRIOT Act (Public Law 107-56) the responsibility for developing and certifying biometric technology standards. NIST
has been supporting biometric standards and evaluation activities for over forty years, starting with automated
fingerprint analysis which began in 1965.
The MITRE Corporation has developed an informative guidance that is widely recognized as the de facto standard
guidance for utilizing JPEG 2000 for the compression of 1000 ppi fingerprint imagery in MTR-04B0000022 [MTR]. While
this document provides a compression profile for 1000 ppi fingerprint imagery using JPEG 2000, the guidance focuses
on lossy compression of fingerprints. Unlike WSQ, JPEG 2000 supports both lossy functionality similar to WSQ (see 1.1),
as well as a lossless mode. This study examines JPEG 2000 in lossless mode and compares its performance to other
lossless compression algorithms.
2. Materials and Methods
A total of 1400 images were selected for this study from the NIST SD-27 special database [SD27]. These images included
various impression types (see Table 2) and included multiple subjects (see Appendix A for more information on the
makeup of the input data set). The images were then processed by the algorithms selected for this study. The
algorithms selected were instrumented specifically for this study to enable the collection of timing data. Effective
compression rate data was calculated from the resulting compressed files at a later time. No specific protocol was put
in place for the ordering of images in the input set, however the images were processed in an arbitrarily stratified
fashion in the order of case numbers 2, 4, 1, 3, 6, 5 and 7.
Each algorithm was tested with all the images independently and the images were processed sequentially without any
delays or throttling of the processing stages.
Table 2 –Impression Types Selected for This Study
Case Number Data Medium Impression Type Count
1 Ink Card Scan Rolled single finger 200
2 Ink Card Scan Flat single finger 200
3 Ink Card Scan Slap-four finger 200
4 Ink Card Scan Latent Lift Image 200
5 Digital Live Scan Rolled single finger 200
6 Digital Live Scan Flat single finger 200
7 Digital Live Scan Slap-four finger 200
Total: 1,400
2.1. Compression Algorithms
The focus of this study is to examine the effectiveness of various lossless compression algorithms on 1000 ppi
fingerprint images. While there are many suitable algorithms, this study focuses on five: two implementations of JPEG
2000 (OpenJPEG and JasPer [JPEG2K]), PNG, RLE used in BMP, and ICER. JPEG 2000 (lossless) and PNG were included in
this selection set because they are currently part of the existing standard for fingerprint data exchange [AN2011].
Although the ANSI/NIST standard also lists lossless JPEG, this format was not examined in this study due to anecdotal
evidence indicating the low rate of acceptance of this compression format for 500 ppi fingerprint images. RLE is one of
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 20 OF 52
the oldest and simplest compression algorithms and was included as a baseline due to its simple algorithm design and
wide implementation availability. ICER is quite new, and was selected to represent a robust state-of-the art approach.
All algorithms were tested on the same machine (See Appendix B for test machine configuration) to eliminate any
computing platform bias.
2.1.1. JPEG 2000
JPEG 2000 is an image compression standard and coding system that was created by the Joint Photographic Experts
Group committee (JPEG) in 2000 to improve on the original JPEG image compression standard’s discrete cosine
transform-based methodology [JPEG] by utilizing a wavelet-based methodology. In addition to providing a lossy
approach, JPEG 2000 also provides for a lossless/reversible filter.
2.1.2. PNG
Portable Network Graphics (PNG) is an image compression standard and coding system that was created by the PNG
Working Group and later accepted as a standard in 2004 (ISO/IEC 15948:2004) [ISO/IEC]. PNG was primarily created to
improve upon and replace GIF (Graphics Interchange Format) as an open image file format not requiring a patent
license. PNG utilizes a 2-stage compression process where a prediction filter is applied to the image in an attempt to
make the data more compressible in the ultimate compression stage. PNG is a very flexible compression algorithm
supporting palette-based color images, grayscale images, and RGB images.
2.1.3. RLE Used in BMP
Run-Length Encoding (RLE) is a very simple form of data compression used by the BMP file format in which sequences
of repetitive data are stored as a single data value with an associated repeat count for that value, rather than as the
original sequence. This compression algorithm is most useful on data that contains many such repeating sequences of
the same value, for example a largely blank image containing only white pixels. RLE is not effective with files that
exhibit any significant entropy4
in structure, and frequently the file size of such data can greatly increase after being
compressed with RLE. This is due to the addition of count values to the byte stream when few data are actually
repeated, resulting in more data added than removed or compressed.
2.1.4. ICER
ICER (Image Compression and Stereo Ranging) [KIELY1],[KIELY2] is a modern wavelet-based image data compression
algorithm specially designed to meet the needs of deep-space exploration applications such as the Mars Exploration
Rover (MER) where state-of-the-art data compression effectiveness and efficiency are primary concerns rather than
catering to general purpose applications. ICER provides lossless and lossy compression modes. For the purposes of this
study it was examined in lossless mode. While ICER is not generally available for commercial use, it has been included as
a reference modern high-end compression algorithm, just as RLE has been included in this study to provide a reference
low-end compression algorithm.
4 Appendix C provides a comprehensive description of entropy as interpreted in this study.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 21 OF 52
2.2. Compression Conditions
Per the investigative goals described in Section 1, this study’s experimental goals can be differentiated into two
overarching tracks each consisting of 6 compression conditions (see Table 3 below). The first track focuses on the
comparison of different compression algorithms (inter-algorithm conditions) while the second track focuses on the
different implementations of the same algorithm (intra-algorithm, JPEG-2000 specifically).
Table 3 - Compression Conditions
Experimental Condition (Inter-Algorithm) Experimental Conditions (Intra-Algorithm)
Non-Compressed OpenJPEG v.1.3, 32-bit
RLE (32-bit) OpenJPEG v.1.3, 64-bit
PNG (32-bit) OpenJPEG v.1.4, 32-bit
JPEG 2000 (OpenJPEG v.1.4, 32-bit) OpenJPEG v.1.4, 64-bit
JPEG 2000 (JasPer 1.900.1, 32-bit) JasPer 1.900.1 32-bit
ICER (32-bit) JasPer 1.900.1 64-bit
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 22 OF 52
3. Analysis
3.1. Performance Metrics
In the case of compression performance measurement, there are three key factors in establishing overall performance.
These factors include the effective compression ratio of the algorithm, the effective throughput of the algorithm when
it compresses the image, and the effective throughput of the algorithm when it decompresses the image.
3.1.1. Effective Compression Ratio Measurement
The effective compression ratio is measured by how much size reduction is obtained relative to the original noncompressed image by applying the algorithm being studied. In examining effective compression across multiple test
images of various sizes and impression type, as was the case in this study, simply comparing resulting image sizes is not
practical as the input images are not of homogenous geometry prior to compression (See Table 30). A simple and
common metric of effective compression is defined by calculating the ratio of the non-compressed (raw) image size to
the compressed image size as defined below:
3.1.2. Effective Compression Throughout
As with other metrics, the effective compression throughput of an algorithm can be measured in many ways such as
including or excluding IO overhead and disk cache effectiveness. For the scope of this study a simple clock-based
approach on a dedicated machine was employed. This is described further in Appendix B. The compression time was
measured by instrumenting each of the compression algorithms to record the difference in the time interval from
launch to when the compressed image has been saved onto a hard disk. No attempt was made to exclude and/or
measure disk IO overhead or to measure any impact of built in disk caching mechanism on compression throughput. All
testing was performed on a single machine to avoid any bias introduced by differences in architecture and
configuration between multiple machines.
3.1.3. Effective Decompression Throughput
The effective decompression throughput of an algorithm was measured similarly to the measurement of effective
compression throughput. The decompression time was measured by instrumenting each of the decompression
algorithms to record the difference in the computer’s real-time-clock from launch to when the image had been
successfully decompressed and saved onto a hard disk. No attempt was made to exclude and/or measure disk IO
overhead or to measure any impact the built in disk caching mechanism has on decompression throughput. All testing
was performed on a single machine to avoid any bias introduced by differences in architecture and configuration
between multiple machines.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 23 OF 52
3.2. Uncertainty Measurement
3.2.1. Normality Test
Examination of the distribution of data in the normal probability plot [CHAMBERS] and the Q-Q Plot on a selected
subset of data from this study (effective compression ratio performance of the PNG algorithm) in Figure 2 and Figure 3
shows that it is not possible to determine the distribution of the data and therefore parametric analysis methods are
not suitable. Because of this we accept the median as the comparison statistic and utilize a bootstrap method which
does not rely on the distribution of the underlying data to assess the uncertainty of the medians estimated from the
observed sample measurements.
Figure 2 - Normal Fit Histogram, Effective
Compression (PNG)
Figure 3 - Q-Q Normality Plot, Effective
Compression (PNG)
3.2.2. Bootstrap Procedure
Given the observed non-normality of the measurements, we employ the sample median of N=1400 measurements of
each compression metric (or N=200 measurements for each of seven different impression types) as our comparison
statistic. We estimate the uncertainty, i.e., standard error, of the median using a bootstrap procedure [WU1], [WU2],
[WU3] wherein we re-compute the median for each of 1,000 samples of size N, sampled randomly with replacement
from the original N measurements. The 95 % confidence interval for the median statistic is then taken as the 0.025th and
0.975th quantiles of the distribution of median replicates. The sample medians and their upper and lower confidence
limits are shown in figures 4 – 9 in sections below.
0
200
400
600
800
1000
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Frequency
Effective Compression Ratio (PNG, N:1)
Histogram
Normal Fit
(Mean=1.7578, SD=0.8641)
-4
-3
-2
-1
0
1
2
3
4
5
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Normal Quantile (Z)
Effective Compression Ratio (PNG, N:1)
Normality Plot (Q-Q)
Normal Fit
(Skewness=1.56, Kurtosis=2.06)
(W = 0.78, p = 0.0000)
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 24 OF 52
3.2.3. Hypothesis Testing
Hypothesis testing is performed using a non-parametric method, specifically the Wilcoxon Signed Rank Test
[WILCOXON], [HOLLANDER]. Differences among the various compression processes are tested for each of three
metrics. We test for differences in each of three metrics between every pair of compression processes without regard
to directionality of differences (two-tailed test). We have two sets of compression conditions (Inter-Algorithm and
Intra-Algorithm described in Section 2.2) under test yielding for each set 15 pairwise comparisons for each of the three
comparison metrics (effective compression rate, compression throughput and decompression throughput). We also
perform the pairwise comparison tests for each of the p = 10 fingerprint groupings including All Data, Latent, Ink Card
Scan All, Ink Card Scan Rolled, Ink Card Scan Flat, Ink Card Scan Slap, Digital Live Scan All, Digital Live Scan Rolled, Digital
Live Scan Flat, and Digital Live Scan Slap. That is, for each pair of compression treatments, we wish to test the null
hypothesis, H0, that the median difference of pairwise measurements is zero. We reject the null hypothesis if we find
the probability of its truth to be less than the Type I error rate, i.e. the probability of incorrectly rejecting the null
hypothesis. Typically, the acceptable Type I error rate is set at 5%, i.e. α = 0.05. If we are able to reject the null
hypothesis, we accept the alternative hypothesis, H1
, that the median of N pairwise differences in measurements is not
equal to zero. Thus, for the present experiments, the null hypotheses, H0, and alternative hypotheses, H1
, for the n=15
pairwise comparisons may be stated mathematically as:
, , , , , ,
, , . , , ,
) 0
) 0
i c k g i d k g
i c k g i d k g
i=1...N;c=1...6;d=1...6(d c;);k=1...3;g=1...9 c,k,g
0
c,k,g
i=1...N;c=1...6;d=1...6(d c);k=1...3;g=1...9 c,k,g
1
c,k,g
H : Med(m - m
H : Med(m - m
where m in the expression denotes a metric further specified by subscripts; i designates a pair of measurements 1…N; c
designates the compression method 1...6; k designates metric 1…3; and g designates fingerprint groupings 1…9. Each
of the paired tests is independent of other tests and the procedure is performed separately for each metric and for
each group of fingerprint images as specified above. Moreover, in a similar analysis, we compare various
implementations of the JPEG 2000 CODEC.
The Wilcoxon Signed Rank test examines differences between pairwise measurements without requiring the
assumption of normality. This test is analogous to the pairwise t-test used to compare pairs of measurements having
distributions known to satisfy the assumptions of normality. Even as subsequent tables and graphs present sample
medians, the hypothesis test comparisons are made between pairs of measurements and not between distributions.
3.3. Algorithmic Comparisons
The first phase of this study entailed comparison of the various algorithms selected for this study and their relative
differences. These differences can be summarized as the differences in the effectiveness of the algorithm in terms of
compressing the image relative to the raw file size, the amount of time needed by the algorithm to compress the image
and the amount of time needed by the algorithm to decompress that image.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 25 OF 52
3.3.1. Effective Compression Rate Comparison
Comparison of median values for effective compression ratios shows that the wavelet-based compression algorithms
(ICER and JPEG 2000) hold a definitive advantage over non-wavelet-based compression algorithms (PNG, RLE) in
compressing mixed image types as shown in Figure 4. The older RLE algorithm in the BMP format appears to be
negatively impacted by the high entropy of fingerprint data where it incurs a negative compression penalty; that is, the
image grows in size after compression relative to the original non-compressed file. One exception emerges however, in
that Digital Live Scan Rolled images achieve higher compression ratios using PNG in contrast to wavelet-based
algorithms. Examination of image entropies shows that Digital Live Scan images typically contain less entropy relative
to Ink Card Scan images, particularly in the case of untextured backgrounds. This renders these images more
compressible using traditional (non-wavelet-based) algorithms.
Figure 4 - Effective Compression Rates (Medians)
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
BMP/RLE (32-bit) 0.5822 0.5793 0.5338 0.5202 0.5333 0.5495 1.2125 1.5448 0.7623 1.1908
Non-Compressed 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
PNG (32-bit) 1.3227 1.3137 1.1492 1.0715 1.1619 1.1928 2.4351 3.0325 1.5901 2.4235
JPEG 2000 (OpenJPEG v.1.4, 32-bit) 1.9551 2.3406 1.6978 1.5099 1.7608 1.7420 2.5555 2.9100 1.8950 2.7331
JPEG 2000 (JasPer 1.900.1, 32-bit) 1.9551 2.3407 1.6978 1.5099 1.7609 1.7420 2.5555 2.9099 1.8951 2.7331
ICER (32-bit) 2.0148 2.4130 1.7459 1.5637 1.8148 1.7921 2.5645 2.9126 1.9343 2.7560
0.5 : 1 1.0 : 1 1.5 : 1 2.0 : 1 2.5 : 1 3.0 : 1 3.5 : 1
All Data
Digital Live Scan - Latent
Ink Card Scan - All
Ink Card Scan - Rolled
Ink Card Scan - Flat
Ink Card Scan - Slap
Digital Live Scan - All
Digital Live Scan - Rolled
Digital Live Scan - Flat
Digital Live Scan - Slap
Effective Compression Rate (Median Ratio)
ICER (32 bit)
JPEG 2000 (JasPer 1.900.1, 32 bit)
JPEG 2000 (OpenJPEG v.1.4, 32 bit)
PNG (32 bit)
Non-Compressed
BMP/RLE (32 bit)
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 26 OF 52
3.3.2. Compression Throughput Comparison
In addition to effective compression ratio, the time an algorithm takes to create a compressed representation of an
image can also be a significant factor in the selection of an algorithm. The most notable example here is mobile
applications where image compression may take place on a device with modest computational resources. It should also
be noted that in common biometric processing workflows, the compression process typically occurs once after image
acquisition, while the image may be decompressed one or more times; therefore the load incurred during compression
is a transient one. Comparison of median values for effective compression time has been provided in Figure 5.
Throughput data shows that while the compression time varies heavily by algorithm implementation, there is some
stratification both by algorithm and image geometry (size). In the case of JPEG 2000, the two implementations of JPEG
2000 (JasPer and OpenJPEG) show significant differences in algorithm compression throughput. Examination of
algorithm throughputs does not show a consistent advantage towards either wavelet or non-wavelet-based
compression algorithms, but does show that compression throughput can vary greatly by specific implementations of
the algorithm.
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
RLE (32-bit) 0.1651 0.1151 0.1329 0.0889 0.0671 0.2848 0.1836 0.1189 0.0737 0.2775
PNG (32-bit) 0.6638 0.9194 0.6733 0.6068 0.2976 3.1707 0.5211 0.5188 0.2956 1.5004
JPEG 2000 (OpenJPEG v.1.4, 32-bit) 1.0878 1.1219 0.9754 0.8274 0.3831 4.0815 1.2043 1.2043 0.4728 3.3437
JPEG 2000 (JasPer 1.900.1, 32-bit) 0.4188 0.4261 0.3916 0.3339 0.1548 1.6176 0.4378 0.4378 0.1764 1.1896
ICER (32-bit) 0.5665 0.5490 0.4098 0.3490 0.1606 1.8771 0.6475 0.6475 0.1888 1.7206
Figure 5 - Compression Throughput (Median Time, seconds)
0.0s 0.5s 1.0s 1.5s 2.0s 2.5s 3.0s 3.5s 4.0s 4.5s
All Data
Digital Live Scan - Latent
Ink Card Scan - All
Ink Card Scan - Rolled
Ink Card Scan - Flat
Ink Card Scan - Slap
Digital Live Scan - All
Digital Live Scan - Rolled
Digital Live Scan - Flat
Digital Live Scan - Slap
Compression Throughput (Median Time, seconds)
ICER (32 bit)
JPEG 2000 (JasPer 1.900.1, 32 bit)
JPEG 2000 (OpenJPEG v.1.4, 32 bit)
PNG (32 bit)
BMP/RLE (32 bit)
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 27 OF 52
3.3.3. Decompression Throughput Comparison
Another important factor in algorithm performance is the amount of time the algorithm takes to decompress the image
for processing and/or viewing. The decompression time can be a significant factor in the selection of an algorithm as it
can occur multiple times in the lifecycle of the biometric record. Comparison of median values for effective
decompression time has been provided in Figure 6. Throughput data shows that while the compression time varies
heavily by algorithm implementation, the non-wavelet-based compression algorithms (PNG, RLE) show an advantage in
decompression throughput. This advantage is most dramatic with PNG. In the case of JPEG 2000, the two
implementations of JPEG 2000 (JasPer and OpenJPEG) show significant differences in algorithm decompression
throughput suggesting differing levels of optimization between specific implementations of the algorithm.
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
RLE (32-bit) 0.2317 0.2680 0.2423 0.2121 0.1147 0.7211 0.1686 0.1479 0.0843 0.4508
PNG (32-bit) 0.0525 0.0566 0.0491 0.0416 0.0364 0.1576 0.0518 0.0509 0.0343 0.1011
JPEG 2000 (OpenJPEG v.1.4, 32-bit) 0.8499 0.8549 0.7269 0.6304 0.2933 3.1451 0.9534 0.9534 0.3414 2.6300
JPEG 2000 (JasPer 1.900.1, 32-bit) 0.3660 0.3649 0.3428 0.2857 0.1323 1.3939 0.3914 0.3914 0.1575 1.0549
ICER (32-bit) 0.9471 0.9294 0.6438 0.5553 0.2794 2.9762 1.1068 1.1068 0.3218 2.8767
Figure 6 - Decompression Throughput (Median Time, seconds
0.0s 0.5s 1.0s 1.5s 2.0s 2.5s 3.0s 3.5s 4.0s 4.5s
All Data
Digital Live Scan - Latent
Ink Card Scan - All
Ink Card Scan - Rolled
Ink Card Scan - Flat
Ink Card Scan - Slap
Digital Live Scan - All
Digital Live Scan - Rolled
Digital Live Scan - Flat
Digital Live Scan - Slap
Decompression Throughput (Median Time, seconds)
ICER (32 bit)
JPEG 2000 (JasPer 1.900.1, 32 bit)
JPEG 2000 (OpenJPEG v.1.4, 32 bit)
PNG (32 bit)
BMP/RLE (32 bit)
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 28 OF 52
3.4. Architecture and Implementation Comparisons
The second phase of this study focused on the JPEG 2000 algorithm and examined the relative differences between the
various implementations and architectural configurations of this algorithm. Again, as with section [3.3] the differences
are summarized as the differences in the effectiveness of the algorithm in terms of compressing the image relative to
the raw (non-compressed) file size, the amount of time needed by the algorithm to compress the image and the
amount of time needed by the algorithm to decompress that image.
3.4.1. Effective Compression Rate Comparison
Comparison of median values for effective compression ratios shows that the various implementations of JPEG 2000
yield nearly identical compression rates.
CODEC Image Type
All Data Ink Card Scan Digital Live Scan
All Latent All Rolled Flat Slap All Rolled Flat Slap
OpenJPEG v.1.3, 32-bit 1.9550 All 1.6978 1.5098 1.7607 1.7420 2.5555 2.9099 1.8950 2.7331
OpenJPEG v.1.3, 64-bit 1.9550 2.3406 1.6978 1.5098 1.7607 1.7420 2.5555 2.9099 1.8950 2.7331
OpenJPEG v.1.4, 32-bit 1.9551 2.3406 1.6978 1.5099 1.7608 1.7420 2.5555 2.9100 1.8950 2.7332
OpenJPEG v.1.4, 64-bit 1.9551 2.3406 1.6978 1.5099 1.7608 1.7420 2.5555 2.9100 1.8950 2.7332
JasPer 1.900.1 32-bit 1.9551 2.3406 1.6978 1.5099 1.7609 1.7420 2.5555 2.9099 1.8951 2.7331
JasPer 1.900.1 64-bit 1.9551 2.3407 1.6978 1.5099 1.7609 1.7420 2.5555 2.9099 1.8951 2.7331
Figure 7 - Effective Compression Rate (Median Ratio, by Platform and Implementation)
0.5 : 1 1.0 : 1 1.5 : 1 2.0 : 1 2.5 : 1 3.0 : 1 3.5 : 1
All Data
Digital Live Scan - Latent
Ink Card Scan - All
Ink Card Scan - Rolled
Ink Card Scan - Flat
Ink Card Scan - Slap
Digital Live Scan - All
Digital Live Scan - Rolled
Digital Live Scan - Flat
Digital Live Scan - Slap
Effective Compression Rate (Median Ratio)
OpenJPEG v.1.3, 32 bit
OpenJPEG v.1.3, 64 bit
OpenJPEG v.1.4, 32 bit
OpenJPEG v.1.4, 64 bit
JasPer 1.900.1 32 bit
JasPer 1.900.1 64 bit
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 29 OF 52
3.4.2. Compression Throughput Comparison
Comparison of median values for effective compression time as shown in Figure 8 shows that platform optimization
and evolution can lead to significant improvements in speed as evident from comparing OpenJPEG v.1.3 and v.1.4.
Architectural differences however are not evident. For example, the JasPer implementation shows a performance
penalty when compiled for a 64-bit architecture, while OpenJPEG appears to have performance gains when operating
in a 64-bit environment.
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
OpenJPEG v.1.3, 32-bit 1.0506 1.0771 0.9126 0.7956 0.3762 3.8532 1.1811 1.1801 0.4614 3.2632
OpenJPEG v.1.3, 64-bit 0.9940 1.0120 0.8658 0.7332 0.3515 3.7440 1.0980 1.0980 0.4266 3.1001
OpenJPEG v.1.4, 32-bit 1.0757 1.0995 0.9282 0.7800 0.3732 4.0248 1.1775 1.1775 0.4527 3.3267
OpenJPEG v.1.4, 64-bit 0.4700 0.4873 0.4042 0.3427 0.1641 1.7316 0.5130 0.5130 0.1904 1.4411
JasPer 1.900.1 32-bit 0.4147 0.4289 0.3748 0.3234 0.1558 1.6068 0.4314 0.4314 0.1752 1.2050
JasPer 1.900.1 64-bit 0.6708 0.7075 0.5928 0.4991 0.2434 2.5116 0.7152 0.7152 0.2846 2.0416
Figure 8 - Compression Throughput (Median Time, seconds, by Platform and Implementation)
0.0s 0.5s 1.0s 1.5s 2.0s 2.5s 3.0s 3.5s 4.0s 4.5s
All Data
Digital Live Scan - Latent
Ink Card Scan - All
Ink Card Scan - Rolled
Ink Card Scan - Flat
Ink Card Scan - Slap
Digital Live Scan - All
Digital Live Scan - Rolled
Digital Live Scan - Flat
Digital Live Scan - Slap
Compression Throughput (Median Time, seconds)
OpenJPEG v.1.3, 32 bit
OpenJPEG v.1.3, 64 bit
OpenJPEG v.1.4, 32 bit
OpenJPEG v.1.4, 64 bit
JasPer 1.900.1 32 bit
JasPer 1.900.1 64 bit
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 30 OF 52
3.4.3. Decompression Throughput Comparison
Comparison of median values for effective decompression time as shown in Figure 9 shows a similar pattern to the data
collected in compression throughput measurements from Figure 8. Again, platform optimization and evolution appear
to lead to significant improvements in decompression speed as evident from comparing OpenJPEG v.1.3 and v.1.4.
Architectural differences do not appear to significantly or consistently impact decompression times. For example, the
JasPer implementation shows a throughput penalty when operating in 64-bit compilation, while OpenJPEG appears to
have performance gains when operating in 64-bit compilation.
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
OpenJPEG v.1.3, 32-bit 0.9203 0.9399 0.7877 0.6785 0.3245 3.4319 1.0434 1.0434 0.3958 2.9190
OpenJPEG v.1.3, 64-bit 0.9029 0.9206 0.7954 0.6707 0.3216 3.4006 0.9984 0.9984 0.3893 2.8150
OpenJPEG v.1.4, 32-bit 0.8578 0.8734 0.7174 0.6083 0.2964 3.1198 0.9494 0.9494 0.3482 2.6870
OpenJPEG v.1.4, 64-bit 0.3587 0.3749 0.3086 0.2617 0.1236 1.3258 0.3930 0.3930 0.1435 1.1019
JasPer 1.900.1 32-bit 0.3631 0.3635 0.3302 0.2871 0.1344 1.3726 0.3859 0.3859 0.1566 1.0729
JasPer 1.900.1 64-bit 0.5728 0.5874 0.4994 0.4320 0.2081 2.1527 0.6179 0.6179 0.2404 1.7299
Figure 9 - Decompression Throughput (Median Time, seconds, by Platform and Implementation)
0.0s 0.5s 1.0s 1.5s 2.0s 2.5s 3.0s 3.5s 4.0s 4.5s
All Data
Digital Live Scan - Latent
Ink Card Scan - All
Ink Card Scan - Rolled
Ink Card Scan - Flat
Ink Card Scan - Slap
Digital Live Scan - All
Digital Live Scan - Rolled
Digital Live Scan - Flat
Digital Live Scan - Slap
Decompression Throughput (Median Time, seconds)
OpenJPEG v.1.3, 32 bit
OpenJPEG v.1.3, 64 bit
OpenJPEG v.1.4, 32 bit
OpenJPEG v.1.4, 64 bit
JasPer 1.900.1 32 bit
JasPer 1.900.1 64 bit
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 31 OF 52
4. Results
4.1. Investigative Goal 1: Examine Effective Compression Rates by Algorithm
One of the primary drivers in the selection of a compression ratio is the effective compression rate obtained using that
algorithm on the data of interest. The first investigative goal of this study is to determine if any particular algorithm
holds a definitive advantage versus the others, and if the measured advantage is statistically significant.
4.1.1. Investigative Analysis 1
Based on the median data (Table 4), ICER provides the best compression rate for mixed image types (“All Data”) with a
median ratio of 2.0148:1 followed by the implementations of JPEG 2000. Pair-wise examination of the data using
Wilcoxon Signed Rank Test with a significance level (alpha) of 0.05 shows that the differences between ICER and the
two JPEG 2000 algorithms are closer than other probabilities to the Type I error threshold of 0.05 for the Digital Live
Scan case, but are still significantly different (Table 5). Hence the null hypothesis is rejected for this comparison as well
as for all other comparisons for which the probability of incorrectly rejecting the null hypothesis are well below the 0.05
level. The data also shows that ICER and JPEG 2000 yield the best compression performance for every case except for
Digital Live Scan Rolled imagery where PNG outperforms all other algorithms.
Table 4 - Effective Compression Rates (Medians), Higher is Better
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
RLE (32-bit) 0.5822 0.5793 0.5338 0.5202 0.5333 0.5495 1.2125 1.5448 0.7623 1.1908
Non-Compressed 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
PNG (32-bit) 1.3227 1.3137 1.1492 1.0715 1.1619 1.1928 2.4351 3.0325 1.5901 2.4235
JPEG 2000 (OpenJPEG v.1.4, 32-bit) 1.9551 2.3406 1.6978 1.5099 1.7608 1.7420 2.5555 2.9100 1.8950 2.7331
JPEG 2000 (JasPer 1.900.1, 32-bit) 1.9551 2.3407 1.6978 1.5099 1.7609 1.7420 2.5555 2.9099 1.8951 2.7331
ICER (32-bit) 2.0148 2.4130 1.7459 1.5637 1.8148 1.7921 2.5645 2.9126 1.9343 2.7560
Table 5 – Effective Compression, Wilcoxon Signed Rank Test, alpha of 0.05
Case Comparison Pair Image Type
All Data Latent Ink Card Scan Digital Live Scan
CODEC 1 CODEC 2 All All All Rolled Flat Slap All Rolled Flat Slap
1 Non-Compressed RLE <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
2 Non-Compressed PNG <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
3 Non-Compressed J2K (OpenJPEG v.1.4) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4 Non-Compressed J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
5 Non-Compressed ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
6 RLE ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
7 RLE J2K (OpenJPEG v.1.4) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
8 RLE J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
9 RLE ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
10 PNG J2K (OpenJPEG v.1.4) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
11 PNG J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
12 PNG ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
13 J2K (OpenJPEG v.1.4) J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0013 <0.0001 <0.0001 <0.0001
14 J2K (OpenJPEG v.1.4) ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0400 <0.0001 <0.0001
15 J2K (JasPer 1.900.1) ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0426 <0.0001 <0.0001
4.1.2. Investigative Result 1
Based on the data, the wavelet algorithms provide consistently better performance than their non-wavelet-based
counterparts with results that are statistically significant. ICER leads the pack, followed by JPEG 2000. The only
exception to this was the Live Scan Rolled image case. It is hypothesized that the lower entropy of this particular image
type may make these images more effectively compressible by PNG. RLE on the other hand seems to suffer a significant
penalty with high entropy images such as Ink Card Scan where the resulting compressed representation is larger than
the original non-compressed images.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 32 OF 52
4.2. Investigative Goal 2: Examine Compression Throughput by Algorithm
Another important driver in the selection of a compression ratio is throughput of the algorithm in terms of the time it
takes to generate the compressed representation of data of interest. Compression typically only occurs once upon the
collection and packaging of the biometric sample and therefore it is not perceived as a source of any significant
recurring load on the processing pathways of the typical biometric system. Compression time however can be a
significant decision driver for devices with modest computational ability such as hand-held or embedded devices where
even a transient load may have very significant impact on the workflow and usability of such devices. The second
investigative goal of this study is to determine if any particular algorithm holds a definitive advantage in terms of
compression throughput, and if the measured advantage is statistically significant.
4.2.1. Investigative Analysis 2
Based on the median data (Table 6), RLE provides the best compression throughput5
for mixed image types (“All
Data”) with a median ratio of 0.1651 seconds, followed by the JasPer implementation of JPEG 2000 and PNG. Pair-wise
examination of the data using Wilcoxon Signed Rank Test with a significance level (alpha) of 0.05 shows that all
comparisons were statistically significant for all comparison pairs and image types, i.e. all probabilities are below the
0.05 level.
Table 6 - Compression Throughput (Median Time, seconds), Lower is Better
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
RLE (32-bit) 0.1651 0.1151 0.1329 0.0889 0.0671 0.2848 0.1836 0.1189 0.0737 0.2775
PNG (32-bit) 0.6638 0.9194 0.6733 0.6068 0.2976 3.1707 0.5211 0.5188 0.2956 1.5004
JPEG 2000 (OpenJPEG v.1.4, 32-bit) 1.0878 1.1219 0.9754 0.8274 0.3831 4.0815 1.2043 1.2043 0.4728 3.3437
JPEG 2000 (JasPer 1.900.1, 32-bit) 0.4188 0.4261 0.3916 0.3339 0.1548 1.6176 0.4378 0.4378 0.1764 1.1896
ICER (32-bit) 0.5665 0.5490 0.4098 0.3490 0.1606 1.8771 0.6475 0.6475 0.1888 1.7206
Table 7 - Compression Throughput, Wilcoxon Signed Rank Test, alpha of 0.05
Case Comparison Pair Image Type
All Data Latent Ink Card Scan Digital Live Scan
CODEC 1 CODEC 2 All All All Rolled Flat Slap All Rolled Flat Slap
1 Non-Compressed RLE <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
2 Non-Compressed PNG <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
3 Non-Compressed J2K (OpenJPEG v.1.4) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4 Non-Compressed J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
5 Non-Compressed ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
6 RLE ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
7 RLE J2K (OpenJPEG v.1.4) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
8 RLE J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
9 RLE ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
10 PNG J2K (OpenJPEG v.1.4) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
11 PNG J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
12 PNG ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
13 J2K (OpenJPEG v.1.4) J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
14 J2K (OpenJPEG v.1.4) ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
15 J2K (JasPer 1.900.1) ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4.2.2. Investigative Result 2
The RLE (using BMP) algorithm provides it with the greatest throughput advantage for the mixed image type case (“All
Data”). The JasPer implementation of JPEG 2000 and ICER follow RLE in throughput performance. It is hypothesized
that the simplicity of the RLE’s compression is most likely responsible for the very high throughput, albeit at the cost of
far less effective compression yield (See 4.1.1). The ordering of throughput is consistent with all image type cases, and
statistically significant for every case.
5
It should be noted that the Non-Compressed case was omitted from this ranking. The Non-Compressed case, if taken in literal
context with the other algorithms, would rank as best in compression throughput due to an elapsed compression time of 0.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 33 OF 52
4.3. Investigative Goal 3: Examine Decompression Throughput by Algorithm
Other than compression ratio, the decompression throughput of an algorithm is the next most important factor in
terms of the time it takes to reconstruct the original image from its compressed representation (decompression).
Image decompression is expected to occur at least one time along the processing pathway of a biometric system, and
possibly many times throughout the lifecycle of that biometric sample; therefore it may be a significant decision driver.
This non-transient load may again be a significant factor with portable devices. The third investigative goal of this study
is to determine if any particular algorithm holds a definitive advantage versus the others in terms of decompression
throughput, and if the measured advantage is statistically significant.
4.3.1. Investigative Analysis 3
Based on the median data (Table 8), PNG provides the best decompression throughput for mixed image types case
(“All Data”) with a median of 0.0525 seconds, followed by RLE and the JasPer implementation of JPEG 2000. Pair-wise
examination of the data using Wilcoxon Signed Rank Test with an alpha of 0.05 shows that all comparisons were
statistically significant for all comparison pairs and image types, with probabilities of a Type I error well below the 0.05
level.
Table 8 - Decompression Throughput (Median Time, seconds)
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
RLE (32-bit) 0.2317 0.2680 0.2423 0.2121 0.1147 0.7211 0.1686 0.1479 0.0843 0.4508
PNG (32-bit) 0.0525 0.0566 0.0491 0.0416 0.0364 0.1576 0.0518 0.0509 0.0343 0.1011
JPEG 2000 (OpenJPEG v.1.4, 32-bit) 0.8499 0.8549 0.7269 0.6304 0.2933 3.1451 0.9534 0.9534 0.3414 2.6300
JPEG 2000 (JasPer 1.900.1, 32-bit) 0.3660 0.3649 0.3428 0.2857 0.1323 1.3939 0.3914 0.3914 0.1575 1.0549
ICER (32-bit) 0.9471 0.9294 0.6438 0.5553 0.2794 2.9762 1.1068 1.1068 0.3218 2.8767
Table 9 - Decompression Throughput, Wilcoxon Signed Rank Test, alpha of 0.05
Case Comparison Pair Image Type
All Data Latent Ink Card Scan Digital Live Scan
CODEC 1 CODEC 2 All All All Rolled Flat Slap All Rolled Flat Slap
1 Non-Compressed RLE <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
2 Non-Compressed PNG <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
3 Non-Compressed J2K (OpenJPEG v.1.4) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4 Non-Compressed J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
5 Non-Compressed ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
6 RLE ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
7 RLE J2K (OpenJPEG v.1.4) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
8 RLE J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
9 RLE ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
10 PNG J2K (OpenJPEG v.1.4) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
11 PNG J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
12 PNG ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
13 J2K (OpenJPEG v.1.4) J2K (JasPer 1.900.1) <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
14 J2K (OpenJPEG v.1.4) ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
15 J2K (JasPer 1.900.1) ICER <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4.3.2. Investigative Result 3
PNG’s advantage in decompression throughput is dramatic and at times an order of magnitude better than the next
closest performer. While PNG’s compression effectiveness may lag behind its wavelet-based counterparts (JPEG 2000
and ICER), the performance advantages in decompression may make it a suitable algorithm for asymmetric use cases
where decompression performance is critical.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 34 OF 52
4.4. Investigative Goal 4: Examine Effective Compression Rate by Impression Type
Another factor in the selection of a compression algorithm is applicability of that algorithm to the taxonomy of the data
that is expected as part of the workflow and processes. If the data has skewed taxonomy (i.e., only Flat images taken
by a Digital Live Scan device) it may be beneficial to select a compression algorithm that performs better given that
data. The fourth investigative goal of this study is to determine if any particular algorithm holds a definitive advantage
versus the others in terms of performance by the type of image processed.
4.4.1. Investigative Analysis 4
The algorithms in this study were ranked in terms of performance for the various criteria of interest (effective
compression ratio, compression throughput and decompression throughput). The results are provided in Table 10
below. Examination of rankings across the columns of Table 10 shows that rankings/stratification of algorithms appear
to be relatively stable and consistent by various image types. There are cases where positions of algorithms with similar
performance rankings switch, but this change in ranking is usually limited to algorithms with adjacent performance
rankings. The notable exception to this is PNG’s effective compression ratio performance with Digital Live Scan Rolled
images where the algorithm jumps from a typical ranking of 4th place to 1st place relative to the other algorithms. For
reference purposes, the non-compressed case has been included in the throughput rankings. The non-compressed
cases here represent the system baseline of zero time required.
Table 10 - Algorithm Performance Rankings (1-Best through 6-Worst)
4.4.2. Investigative Result 4
Across all image types, the rankings for the various measurement criteria appear to be stratified in a stable fashion with
very few changes in rankings6
, and those changes in rankings are typically only a one position change in ranking (i.e., in
terms of effective compression ratio RLE ranks 5
th on Digital Live Scan Slap but drops to 6
th for Digital Live Scan Flat).
The one exception is PNG for Digital Live Scan Rolled images where PNG jumps from 4th place ranking for all other
image types to 1st place ranking.
6 When examining the ranking data in this section, it would be prudent to examine the actual values of the metric in
question as a change in rank order may be the result of a very small change in the underlying metric.
Factor CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
Effective Ratio
Non-Compressed 5 5 5 5 5 5 6 6 5 6
BMP/RLE (32 bit) 6 6 6 6 6 6 5 5 6 5
PNG (32 bit) 4 4 4 4 4 4 4 1 4 4
JPEG 2000 (OpenJPEG v.1.4, 32 bit) 3 3 3 3 3 3 2 3 3 2
JPEG 2000 (JasPer 1.900.1, 32 bit) 2 2 2 2 2 2 3 4 2 3
ICER (32 bit) 1 1 1 1 1 1 1 2 1 1
Compression Throughput
Non-Compressed 1 1 1 1 1 1 1 1 1 1
BMP/RLE (32 bit) 2 2 2 2 2 2 2 2 2 2
PNG (32 bit) 5 5 5 5 5 5 4 4 5 4
JPEG 2000 (OpenJPEG v.1.4, 32 bit) 6 6 6 6 6 6 6 6 6 6
JPEG 2000 (JasPer 1.900.1, 32 bit) 3 3 3 3 3 3 3 3 3 3
ICER (32 bit) 4 4 4 4 4 4 5 5 4 5
Decompression Throughput
Non-Compressed 1 1 1 1 1 1 1 1 1 1
BMP/RLE (32 bit) 3 3 3 3 3 3 3 3 3 3
PNG (32 bit) 2 2 2 2 2 2 2 2 2 2
JPEG 2000 (OpenJPEG v.1.4, 32 bit) 5 5 6 6 6 6 5 5 6 5
JPEG 2000 (JasPer 1.900.1, 32 bit) 4 4 4 4 4 4 4 4 4 4
ICER (32 bit) 6 6 5 5 5 5 6 6 5 6
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 35 OF 52
4.5. Investigative Goal 5: Comparison of JPEG 2000 Implementation With Respect to Effective
Compression Rate
With certain algorithms, performance of the algorithm can vary greatly by the specific implementation of that
algorithm (such as different vendors producing compatible but different algorithms), or the architecture for which the
algorithm was built (such as 32-bit vs. 64-bit). The fifth investigative goal of this study is to see if such implementation
differences do indeed introduce significant operational differences in terms of effective compression ratio. Since the
focus of this study is the JPEG 2000 algorithm, this examination was limited to lossless JPEG 2000 and does not include
RLE, ICER or PNG.
4.5.1. Investigative Analysis 5
The median compression rate data in Table 11 shows that the various implementations of JPEG 2000 performed nearly
identically. Pair-wise examination of the data using the Wilcoxon Signed Rank Test with an alpha of 0.05 however
shows significant differences at values which appear to be identical. Further examination of the statistical processes
used have lead the team to conclude that pairwise analysis of real numbers may not be ideal for comparisons of nearly
identical real values where subtle but operationally insignificant differences can lead to statistically significant
differences between two such lists of values. A further discussion of this effect is provided in Section 6 of this report.
Table 11 - Effective Compression Rates (Medians) by Implementation, Higher is Better
CODEC All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
OpenJPEG v.1.3, 32-bit 1.9550 2.3406 1.6978 1.5098 1.7607 1.7420 2.5555 2.9099 1.8950 2.7331
OpenJPEG v.1.3, 64-bit 1.9550 2.3406 1.6978 1.5098 1.7607 1.7420 2.5555 2.9099 1.8950 2.7331
OpenJPEG v.1.4, 32-bit 1.9551 2.3406 1.6978 1.5099 1.7608 1.7420 2.5555 2.9100 1.8950 2.7332
OpenJPEG v.1.4, 64-bit 1.9551 2.3406 1.6978 1.5099 1.7608 1.7420 2.5555 2.9100 1.8950 2.7332
JasPer 1.900.1 32-bit 1.9551 2.3407 1.6978 1.5099 1.7609 1.7420 2.5555 2.9099 1.8951 2.7331
JasPer 1.900.1 64-bit 1.9551 2.3407 1.6978 1.5099 1.7609 1.7420 2.5555 2.9099 1.8951 2.7331
Table 12 - Effective Compression by Implementation, Wilcoxon Signed Rank Test, at alpha = 0.05
Case Comparison Pair Image Type
All Data Latent Ink Card Scan Digital Live Scan
CODEC 1 CODEC 2 All All All Rolled Flat Slap All Rolled Flat Slap
1 OpenJPEG v.1.3, 32-bit OpenJPEG v.1.3, 64-bit 0.3173 1.0000 0.3173 1.0000 0.3173 1.0000 1.0000 1.0000 1.0000 1.0000
2 OpenJPEG v.1.3, 32-bit OpenJPEG v.1.4, 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
3 OpenJPEG v.1.3, 32-bit OpenJPEG v.1.4, 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4 OpenJPEG v.1.3, 32-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0001 <0.0001 <0.0001
5 OpenJPEG v.1.3, 32-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0001 <0.0001 <0.0001
6 OpenJPEG v.1.3, 64-bit OpenJPEG v.1.4, 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
7 OpenJPEG v.1.3, 64-bit OpenJPEG v.1.4, 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
8 OpenJPEG v.1.3, 64-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0001 <0.0001 <0.0001
9 OpenJPEG v.1.3, 64-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0001 <0.0001 <0.0001
10 OpenJPEG v.1.4, 32-bit OpenJPEG v.1.4, 64-bit 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
11 OpenJPEG v.1.4, 32-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0012 <0.0001 <0.0001 <0.0001
12 OpenJPEG v.1.4, 32-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0012 <0.0001 <0.0001 <0.0001
13 OpenJPEG v.1.4, 64-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0012 <0.0001 <0.0001 <0.0001
14 OpenJPEG v.1.4, 64-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0012 <0.0001 <0.0001 <0.0001
15 JasPer 1.900.1 32-bit JasPer 1.900.1 64-bit 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
4.5.2. Investigative Result 5
Examination of various implementations of JPEG 2000 in both 32-bit and 64-bit architectures yielded nearly identical
results in effective compression rates therefore architectural differences are not expected to be a factor in the
adoption of a particular implementation of lossless JPEG 2000 CODECs.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 36 OF 52
4.6. Investigative Goal 6: Comparison of JPEG 2000 Implementation With Respect to
Compression Throughput
With certain algorithms, throughput of the algorithm can also vary greatly by the specific implementation of that
algorithm (such as different vendors producing compatible but different algorithms), or the architecture for which the
algorithm was built (such as 32-bit vs. 64-bit). The sixth investigative goal of this study is to see if such implementation
differences do indeed introduce significant operational differences in terms of compression throughput. As with the
previous section, since the focus of this study is the JPEG 2000 algorithm, this investigative goal was limited to lossless
JPEG 2000 and does not include RLE, ICER or PNG.
4.6.1. Investigative Analysis 6
The median compression throughput data in Table 13 shows that there are differences evident between the various
implementations of JPEG 2000. Pair-wise examination of the data using Wilcoxon Signed Rank Test with an alpha of
0.05 also confirmed significant differences between the OpenJPEG 32-bit v1.3 and 32-bit v1.4 implementations for all
cases except Ink Scan Flat, Digital Live Scan All and Digital Live Scan Rolled.
Table 13 - Compression Throughput (Median Time, seconds) by Implementation, Lower is Better
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
OpenJPEG v.1.3, 32-bit 1.0506 1.0771 0.9126 0.7956 0.3762 3.8532 1.1811 1.1801 0.4614 3.2632
OpenJPEG v.1.3, 64-bit 0.9940 1.0120 0.8658 0.7332 0.3515 3.7440 1.0980 1.0980 0.4266 3.1001
OpenJPEG v.1.4, 32-bit 1.0757 1.0995 0.9282 0.7800 0.3732 4.0248 1.1775 1.1775 0.4527 3.3267
OpenJPEG v.1.4, 64-bit 0.4700 0.4873 0.4042 0.3427 0.1641 1.7316 0.5130 0.5130 0.1904 1.4411
JasPer 1.900.1 32-bit 0.4147 0.4289 0.3748 0.3234 0.1558 1.6068 0.4314 0.4314 0.1752 1.2050
JasPer 1.900.1 64-bit 0.6708 0.7075 0.5928 0.4991 0.2434 2.5116 0.7152 0.7152 0.2846 2.0416
Table 14 - Compression Throughput by Implementation, Wilcoxon Signed Rank Test, at alpha = 0.0033
Case Comparison Pair Image Type
All Data Latent Ink Card Scan Digital Live Scan
CODEC 1 CODEC 2 All All All Rolled Flat Slap All Rolled Flat Slap
1 OpenJPEG v.1.3, 32-bit OpenJPEG v.1.3, 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
2 OpenJPEG v.1.3, 32-bit OpenJPEG v.1.4, 32-bit <0.0001 <0.0001 <0.0001 <0.0001 0.0758 <0.0001 0.4226 0.2151 <0.0001 <0.0001
3 OpenJPEG v.1.3, 32-bit OpenJPEG v.1.4, 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4 OpenJPEG v.1.3, 32-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
5 OpenJPEG v.1.3, 32-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
6 OpenJPEG v.1.3, 64-bit OpenJPEG v.1.4, 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
7 OpenJPEG v.1.3, 64-bit OpenJPEG v.1.4, 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
8 OpenJPEG v.1.3, 64-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
9 OpenJPEG v.1.3, 64-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
10 OpenJPEG v.1.4, 32-bit OpenJPEG v.1.4, 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
11 OpenJPEG v.1.4, 32-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
12 OpenJPEG v.1.4, 32-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
13 OpenJPEG v.1.4, 64-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
14 OpenJPEG v.1.4, 64-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
15 JasPer 1.900.1 32-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4.6.2. Investigative Result 6
Examination of various implementations and architectures of JPEG 2000 yielded statistically significant differences in
compression throughput between OpenJPEG 32-bit v1.3 and 32-bit v1.4 implementations in all but three cases: Ink Scan
Flat, Digital Live Scan All and Digital Live Scan Rolled. Leading the pack is the highly optimized 32-bit implementation of
the JasPer CODEC, followed by the 64-bit implementation of the OpenJPEG v.1.4 CODEC.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 37 OF 52
4.7. Investigative Goal 7: Comparison of JPEG 2000 Implementation With Respect to
Decompression Throughput
As with compression throughput, the performance of algorithms can vary for the decompression process by
implementation (i.e., same algorithm but different vendors or system architecture). The seventh investigative goal of
this study is to see if such implementation differences do indeed introduce significant operational differences in terms
of decompression throughput. Once again since the focus of this study is the JPEG 2000 algorithm, this investigative
goal was limited to lossless JPEG 2000 and does not include RLE, ICER or PNG.
4.7.1. Investigative Analysis 7
The median decompression throughput data in Table 15 shows that there are differences evident between the various
implementations of JPEG 2000. Pair-wise examination of the data using Wilcoxon Signed Rank Test with a significance
level of 0.05 also confirmed significant differences for all of these cases but two (OpenJPEG 32-bit v1.3 vs. 32-bit v1.4
with Ink Card Scan Rolled, and OpenJPEG 32-bit v1.4 vs. JasPer 32-bit with Latent imagery).
Table 15 - Decompression Throughput (Median Time, seconds) by Implementation, Lower is Better
CODEC Image Type
All Data Latent Ink Card Scan Digital Live Scan
All All All Rolled Flat Slap All Rolled Flat Slap
OpenJPEG v.1.3, 32-bit 0.9203 0.9399 0.7877 0.6785 0.3245 3.4319 1.0434 1.0434 0.3958 2.9190
OpenJPEG v.1.3, 64-bit 0.9029 0.9206 0.7954 0.6707 0.3216 3.4006 0.9984 0.9984 0.3893 2.8150
OpenJPEG v.1.4, 32-bit 0.8578 0.8734 0.7174 0.6083 0.2964 3.1198 0.9494 0.9494 0.3482 2.6870
OpenJPEG v.1.4, 64-bit 0.3587 0.3749 0.3086 0.2617 0.1236 1.3258 0.3930 0.3930 0.1435 1.1019
JasPer 1.900.1 32-bit 0.3631 0.3635 0.3302 0.2871 0.1344 1.3726 0.3859 0.3859 0.1566 1.0729
JasPer 1.900.1 64-bit 0.5728 0.5874 0.4994 0.4320 0.2081 2.1527 0.6179 0.6179 0.2404 1.7299
Table 16 - Decompression Throughput, Wilcoxon Signed Rank Test, alpha of 0.05
Case Comparison Pair Image Type
All Data Latent Ink Card Scan Digital Live Scan
CODEC 1 CODEC 2 All All All Rolled Flat Slap All Rolled Flat Slap
1 OpenJPEG v.1.3, 32-bit OpenJPEG v.1.3, 64-bit <0.0001 <0.0001 <0.0001 0.4836 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
2 OpenJPEG v.1.3, 32-bit OpenJPEG v.1.4, 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
3 OpenJPEG v.1.3, 32-bit OpenJPEG v.1.4, 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4 OpenJPEG v.1.3, 32-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
5 OpenJPEG v.1.3, 32-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
6 OpenJPEG v.1.3, 64-bit OpenJPEG v.1.4, 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
7 OpenJPEG v.1.3, 64-bit OpenJPEG v.1.4, 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
8 OpenJPEG v.1.3, 64-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
9 OpenJPEG v.1.3, 64-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
10 OpenJPEG v.1.4, 32-bit OpenJPEG v.1.4, 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
11 OpenJPEG v.1.4, 32-bit JasPer 1.900.1 32-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
12 OpenJPEG v.1.4, 32-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
13 OpenJPEG v.1.4, 64-bit JasPer 1.900.1 32-bit <0.0001 0.2684 <0.0001 <0.0001 <0.0001 <0.0001 0.00029 <0.0001 <0.0001 <0.0001
14 OpenJPEG v.1.4, 64-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
15 JasPer 1.900.1 32-bit JasPer 1.900.1 64-bit <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
4.7.2. Investigative Result 7
Examination of various implementations of JPEG 2000 in both 32-bit and 64-bit architectures yielded statistically
significant differences in decompression throughput in all but two cases. These were OpenJPEG 32-bit (v1.3 and v1.4)
on Ink Card Scan Rolled and Digital Live Scan Rolled, and OpenJPEG 32-bit v1.4 and JasPer 32-bit on Latent. Leading the
pack is the 64-bit implementation of the OpenJPEG v.1.4 followed by the 32-bit implementation of the JasPer CODEC. It
should also be noted that for the same algorithm version, the 32-bit compilation of that version provided lower
decompression throughput than the 64-bit in almost all the cases, except with the JasPer implementation which favors
the 64-bit platform.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 38 OF 52
4.8. Investigative Goal 8: Comparison of Implementation Complexity
Another factor that can impact the selection of an algorithm is the complexity of the code relative to the target
platform being used. A true complexity analysis requires profiling of the algorithm with instrumented source code
during typical execution. A rudimentary method of complexity analysis can be an estimation of complexity based on the
lines of code in the source files that comprise the algorithm implementation. This complexity estimation is anecdotal
and is not considered a good measure of complexity as a few lines of code may be executed at an exponential
redundancy while many lines of code may execute in linear fashion. This section has been included nonetheless for
discussion purposes.
4.8.1. Investigative Analysis 8
The source files of each respective algorithm were processed; the number of lines of code (LOC), the number of
comments/empty line breaks, and the number of individual source files were all measured. The results of these
measures are presented below in Table 17.
Table 17 - Implementation Complexity
CODEC Gross Lines of
Code (LOC)
Lines of
Comments/Blanks
Net Lines of Code
(LOC)
Number of
Source Files
ICER 6883 991 5892 19
RLE used in BMP 39812 1713 38099 42
JasPer 42235 3857 38378 108
OpenJPG 1.3 70076 11601 58475 163
PNG (LIBPNG+ZLIB) 78893 9094 69799 116
OpenJPG 1.4 87676 14257 73419 179
4.8.2. Investigative Result 8
Based on the LOC analysis, the highly optimized ICER CODEC provides the lowest LOC count, and may therefore be the
least complex. Based on the intended purpose of ICER being run on an autonomous space craft with limited
computational resources, this finding would agree with that goal as ICER’s codebase appears to be a fraction of even
the second least complex algorithm, RLE.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 39 OF 52
5. Conclusions
This experiment was conducted with eight primary goals:
1. Examine Effective Compression Rates by Algorithm
2. Examine Compression Throughput by Algorithm
3. Examine Decompression Throughput by Algorithm
4. Examine Effective Compression Rate by Impression Type
5. Compare JPEG 2000 Implementation With Respect to Effective Compression Rate
6. Compare JPEG 2000 Implementation With Respect to Compression Throughput
7. Compare JPEG 2000 Implementation With Respect to Decompression Throughput
8. Compare Implementation Complexity
Based on the results, it can be concluded that the wavelet-based algorithms (JPEG 2000 and ICER) provide far better
effective compression rates than their non-wavelet-based counterparts (PNG and RLE using BMP) on mixed-image data
consisting of Ink Card Scan, Digital Live Scan, and Latent. Given specific impression types, the PNG algorithm leads all
others for the case of Digital Live Scan Rolled. It is hypothesized that the lower entropy of this particular image type
may make these images more effectively compressible by PNG.
In terms of compression throughput, the non-wavelet-based algorithms lead the pack with the exception of the highly
optimized JasPer 2000 (32-bit) which provides compression throughputs on par with the non-wavelet-based
algorithms. With decompression however, PNG clearly leads the pack by a large margin. While PNG has many uses, it
was primarily designed to be a format suitable for the transmission and display of raster images transmitted across
networks and displayed in web browsers. This is a highly asymmetric use case where images are typically compressed
once and viewed/decompressed many times. Optimization for such use cases, (compress once, view many times), as
well as algorithm complexity, have provided PNG this edge in decompression throughput.
In examining algorithm performance with respect to the various image/impression types used in this study, the image
type used did not appear to have a large impact on changing the rank order dramatically. Where a ranking shift was
noted, this shift was typically only one rank position (i.e., an algorithm that ranked 6th best for Ink Card Scan Rolled
imagery may have fared slightly better in the 5th place with another image type for a given measurement criteria). The
one exception was PNG, which jumped from a 4
th place ranking with All Data in terms of effective compression rate to
1
st place for Digital Live Scan Rolled. Considering its fast decompression times, the PNG algorithm may be particularly
well suited for Digital Live Scan Rolled images.
In studying the behavioral differences of various platforms or implementations of a given algorithm, examination of
various implementations of JPEG 2000 in both 32-bit and 64-bit architectures yielded nearly identical results in terms of
effective compression rates with some very small differences. While these small differences were statistically significant
in almost all cases, they are not operationally relevant (for example, where OpenJPEG v.1.3 yields an effective rate of
1.9550:1 and OpenJPEG v.1.4 yields an effective rate of 1.9551:1). The various architectures did exhibit statistically
significant differences in throughput but these differences were not consistent enough to make a generalized
conclusion. For example, in almost all cases, 32-bit OpenJPEG implementation was slower than the 64-bit OpenJPEG
implementation. For all those cases, the reverse was true with JasPer where the 32-bit version excelled.
Finally, based on a simple comparison of the number of lines of code that comprised each algorithm, the ICER CODEC is
comprised of the fewest lines of code, and may perhaps be the least complex. The next two algorithms are RLE and
JasPer.
Based on this study it can be generally concluded (with specific exceptions as noted) that the non-wavelet-based
algorithms provide an edge in throughput while the wavelet-based algorithms provide an edge in effective
compression rates.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 40 OF 52
6. Discussions and Future Work
Throughout this study, most of the selected measures of performance yielded real (floating point) numbers of very
high precision. Comparison of such numbers in cases of very similar (but different) value may yield differences which
can be operationally inconsequential, but would nonetheless be statistically significant. In this study, we attempted to
mitigate this by conducting preliminary mathematical operations at six significant digit precision with the final stages of
analysis at a precision of four. Such a strategy may only be effective in the scope of image sizes used in this study and
generalization of it may call for adjustment of the rounding depending on the distribution of expected image sizes. This
may be a ripe topic for future study where the very metrics of compression parameters can be explored in a framework
relevant to the systems for which the operations are intended.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 41 OF 52
References
Publications and Reports
BRADLEY1 Bradley, Jonathan N., Brislawn, Christopher M., Hopper, Thomas. (1993). “FBI wavelet/scalar
quantization standard for grayscale fingerprint image compression”, SPIE Conference on Visual
Information Processing II, eds. Huck, Friedrich O., Juday, Richard D. pp. 293-304.
BRADLEY2 Bradley, J. N., Brislawn, C. M. (1994). “The wavelet/scalar quantization compression standard for
digital fingerprint images”. IEEE International Symposium on Circuits and Systems, 1994. ISCAS '94., 1994.
Vol 3. pp. 205-208.
BRISLAWN1 Brislawn, Christopher M. (1996). “Wavelet Scalar Quantization Compression Standard for Fingerprint
Images”. Proceeding Conference on Signal Image Processing and Applications.
BRISLAWN2 Brislawn, Christopher M. (2002). The FBI Fingerprint Image Compression Standard.
http://wwwc3.lanl.gov/~brislawn/FBI/FBI.html, updated 25 June 2001 (accessed 01/31/2014).
CHAMBERS Chambers, John; William Cleveland, Beat Kleiner, and Paul Tukey (1983). Graphical Methods for Data
Analysis. Wadsworth
CIA “The World Factbook”, https://www.cia.gov/library/publications/the-world-factbook/geos/us.html,
Retrieved 2010-12-06.
CURRANEVERETT
Curran-Everett, Douglas (2000). “Multiple comparisons: philosophies and illustrations”, American
Journal of Physiology - Regulatory, Integrative and Comparative Physiology, 279, R1 – R8.
DOJ "The Science of Fingerprints" [rev. December 1984], United States Department of Justice, Federal
Bureau of Investigation, USGPO, ISBN 0-16-050541-0, Page 18.
EYPASCH Eypasch, Ernst; Rolf Lefering, C K Kum, Hans Troidl (1995-09-02). "Probability of adverse events that
have not yet occurred: a statistical reminder". BMJ 311 (7005): 619-620. PMID 7663258. PMC 2550668.
http://www.bmj.com/cgi/content/full/311/7005/619. Retrieved 2008-04-15.
FITZPATRICK Fitzpatrick, M. et al. 1994, “WSQ Compression / Decompression Algorithm Test Report”, IAI Annual
Conference.
GALTON Galton, F. (2005). Finger prints. Mineola, NY: Dover Publications. (Original work published 1892)
HANLEY Hanley, JA; Lippman-Hand A (1983). "If nothing goes wrong, is everything alright?". JAMA 249 (13):
1743-5. PMID 6827763.
HOLLANDER Hollander, M., Wolfe, D. A. (1999). Non-Parametric Statistical Methods, 2nd Ed, John Wiley &
Sons:NY, 787 pages.
HOPPER Hopper, T., Preston, F. (1992). “Compression of grey-scale fingerprint images”, Proceedings of Data
Compression Conference, 1992. DCC '92. pages 309-318.
JAIN Jain, A., “Pores and Ridges: High-Resolution Fingerprint Matching Using Level 3 Features”, IEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol. 29, No. 1, January 2007.
JOVANOVIV Jovanoviv, B. D., and P. S. Levy. A look at the Rule of Three. The American Statistician, Vol. 51, No. 2,
May 1997, pp. 137-139.
KIELY1 Kiely, A. and Klimesh, M. “The ICER Progressive Wavelet Image Compressor,” The Interplanetary
Network Progress Report 42-155, July–September 2003, Jet Propulsion Laboratory, Pasadena,
California, pp. 1–46, November 15, 2003. http://ipnpr.jpl.nasa.gov/progress report/42-155/155J.pdf
KIELY2 Kiely, A. and Klimesh, M. “Preliminary Image Compression Results from the Mars Exploration
Rovers,” The Interplanetary Network Progress Report, vol. 42-156, Jet Propulsion Laboratory,
Pasadena, California, pp. 1–8, February 15, 2004. http://ipnpr.jpl.nasa.gov/progress report/42-
156/156I.pdf
LIBERT “A 1D Spectral Image Validation/Verification Metric for Fingerprints”. Libert, J.M.; Grantham, J.;
Orandi, S. August 19, 2009. http://www.nist.gov/customcf/get_pdf.cfm?pub_id=903078. Retrieved
2011-01-12.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 42 OF 52
LIKERT Likert, R. (1932). A Technique for the Measurement of Attitudes, Archives of Psychology 140, 55.
MTR "Profile for 1000 ppi Fingerprint Compression". Lepley, M.A.
http://www.mitre.org/work/tech_papers/tech_papers_04/lepley_fingerprint/lepley_fingerprint.pdf.
Retrieved 2011-01-11.
NIST1 National Institute of Standards and Technology. Summary of NIST Patriot Act Recommendations.
Gaithersburg, MD. Retrieved January 4, 2007 from
http://www.itl.nist.gov/iad/894.03/pact/NIST_PACT_REC.pdf
NIST2 "NIST Biometric Image Software". http://Fingerprint.nist.gov/NFIS/. Retrieved 2011-01-12.
NISTIR7778 Orandi, S., Libert, J. M., Grantham, J. D., Ko, K., Wood, S.S., Wu, J. Effects of JPEG 2000 Image
Compression on 1000 ppi Fingerprint Imagery, NIST Interagency Report 7778, National Institutes of
Standards and Technology, Gaithersburg, MD. April 11, 2011, 72 pages.
NISTIR7781 Libert, J. M., Orandi, S., Grantham, J. D. Comparison of the WSQ and JPEG 2000 Image Compression
Algorithms on 500 ppi Fingerprint Imagery, NIST Interagency Report 7781, National Institutes of
Standards and Technology, Gaithersburg, MD March 2012, 56 pages.
OPENJPEG "OpenJPEG library : an open source JPEG 2000 codec".
http://www.OpenJPEG.org/index.php?menu=news. Retrieved 2011-01-12.
SD27 M.D. Garris & R.M. McCabe, "NIST Special Database 27: Fingerprint Minutiae from Latent and
Matching Tenprint Images," NIST Technical Report NISTIR 6534 & CD-ROM, June 2000.
SHAPIRO Shapiro, S. S.; Wilk, M. B. (1965). "An analysis of variance test for normality (complete samples)".
Biometrika 52 (3-4): 591–611. doi:10.1093/biomet/52.3-4.591. JSTOR 2333709MR205384.
WILCOXON Wilcoxon, Frank (Dec 1945). "Individual comparisons by ranking methods". Biometrics Bulletin 1 (6):
80–83.
WRIGHT Wright, Paul, S. P-values for simultaneous inference. Biometrics, Vol. 48, No. 4 (Dec. 1992), pp. 1005-
1013.
WU1 Wu, Jin Chu, Alvin F. Martin, and Raghu N. Kacker. Measures, uncertainties, and significance test in
operational ROC analysis, Journal of Research of the National Institute of Standards and Technology,
116(1), 517-537, (2011).
WU2 Wu, Jin Chu. Studies of Operational Measurement of ROC Curve on Large Fingerprint Data Sets Using
Two-Sample Bootstrap. NISTIR 7449, U.S. Department of Commerce, National Institute of Standards
and Technology, September 2007, 25 pages.
WU3 Wu, Jin Chu. Operational Measures and Accuracies of ROC Curve on Large Fingerprint Data Sets.
NISTIR 7495, U.S. Department of Commerce, National Institute of Standards and Technology, May
2008, 23 pages.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 43 OF 52
Standards
AN2011 NIST Special Publication 500-290: American National Standard for Information Systems : Data Format
for the Interchange of Fingerprint, Facial & Other Biometric Information (ANSI/NIST-ITL 1-2011).
Approved November 2011.
ISO/IEC ISO/IEC 15948:2004 – Information technology - Computer graphics and image processing -- Portable
Network Graphics (PNG): Functional specification.
http://www.iso.org/iso/iso_catalogue/catalogue_tc/catalogue_detail.htm?csnumber=29581 Retrieved
2012-01-11.
JPEG "T.81 : Information technology – Digital compression and coding of continuous-tone still images –
Requirements and guidelines". http://www.itu.int/rec/T-REC-T.81. Retrieved 2011-01-12.
JPEG2K "ISO/IEC 15444-1:2004 - Information technology -- JPEG 2000 image coding system: Core coding system".
http://www.iso.org/iso/iso_catalogue/catalogue_ics/catalogue_detail_ics.htm?csnumber=27687.
Retrieved 2009-11-01.
WSQ "WSQ Gray-Scale Fingerprint Image Compression Specification" Version 3.1.
https://www.fbibiospecs.org/docs/WSQ_Gray-scale_Specification_Version_3_1.pdf. Retrieved 2010-01-11.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 44 OF 52
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 45 OF 52
Appendix A. Dataset7 Makeup
For the Ink Card Scan portion of the tests, the study utilized fingerprint images based on the Base Demonstration
Model (BDM) fingerprints utilized in early tests of the FBI IAFIS system, and later used as the basis for the NIST SD-27
special database [SD27]. This Ink Card Scan data was collected as a result of law enforcement activities and represents
actual field data with collection dates ranging from 08/18/1973 through 04/12/1994. The original FD-249 fingerprint
collection cards with these images were retrieved by NIST and rescanned at 1000 ppi by NIST personnel under
controlled conditions. The images were scanned at 8 bits per pixel gray-scale using FBI certified software (Appendix F
compliant) and stored in a non-compressed format to ensure no compression anomalies were introduced into the
original set.
For the Digital Live Scan portion of the tests, the study again utilized actual operational data captured during normal
enforcement activities with collection dates ranging from 01/04/2010 through 04/13/2010. The Digital Live Scan data was
stored so as never to have been subject to lossy compression.
Where possible, the image sets were equally balanced by gender, finger, pattern class and hand. It should be noted that
balancing equally does not follow the natural demographic behavior of the population such as gender (48 % males/52 %
female [CIA]) or pattern class (65 % Loops, 30 % Whorls, 5 % Arches [DOJ]). The goal in having equal distributions was to
avoid the potential statistical bias of very small subsamples. That is, all subsamples were equally important with respect
to compression irrespective of their relative incidence in the population.
Demographic Make-up of Ink Card Scan Datasets
Ink Card Scan images used in this study consisted of 200 each of Rolled, Flat, and Slap.
Table 18 - Ink Card Scan Data classification by Impression Type
All Data
Impression Type Males Females Right Left
Flat Single Finger 100 100 96 104
Rolled Single Finger 100 100 96 104
Four Finger Slaps 100 100 100 100
Table 19 - Ink Card Scan Pattern Classification for Single Finger Images by Impression Type
Data From Females (Single Finger) Data From Males (Single Finger)
Pattern Class Flat Rolled Right Left Pattern Class Flat Rolled Right Left
Arch 34 34 30 38 Arch 34 34 34 34
Loop 33 33 32 34 Loop 33 33 32 34
Whorl 33 33 34 32 Whorl 33 33 34 32
Total 100 100 96 104 Total 100 100 100 100
7
This dataset was chosen to meet constraints required by the NISTIR 7778, but this does not lessen its applicability to this study.
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 46 OF 52
Table 20 - Ink Card Scan Pattern Classification for Single Finger Images by Finger (Females)
Data From Females (Single Finger)
Pattern Class
R. Thumb
R. Index
R. Middle
R. Ring
R. Little
L. Thumb
L. Index
L. Middle
L. Ring
L. Little
Total
Arch 4 4 3 4 0 3 5 4 3 4 34
Loop 3 3 4 3 3 4 3 3 4 3 33
Whorl 3 4 3 3 4 3 3 4 3 3 33
Total 10 11 10 10 7 10 11 11 10 10 100
Table 21 - Ink Card Scan Pattern Classification for Single Finger Images by Finger (Males)
Data From Males (Single Finger)
Pattern Class
R. Thumb
R. Index
R. Middle
R. Ring
R. Little
L. Thumb
L. Index
L. Middle
L. Ring
L. Little
Total
Arch 4 3 3 4 3 3 4 3 3 4 34
Loop 3 3 4 3 3 4 3 3 4 3 33
Whorl 3 4 3 3 4 3 3 4 3 3 33
Total 10 10 10 10 10 10 10 10 10 10 100
Make-up of the digital live Scan data sets
Digital live Scan images used in this study consisted of 200 Rolled,, 200 Flat, and 200 Four Finger Slap impressions.
Table 22 – Digital Live Scan Data Classification by Impression Type
All Data Records
Impression Type Males Females Right Left
Flat Single Finger 100 100 100 100
Rolled Single Finger 100 100 100 100
Four Finger Slaps 100 100 96 104
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 47 OF 52
Table 23 – Digital Live Scan Pattern Classification for Single Finger Images by Impression Type
Data From Females (Single Finger) Data From Males (Single Finger)
Pattern Class Flat Rolled Right Left Pattern Class Flat Rolled Right Left
Arch 33 33 36 30 Arch 33 33 36 30
Loop 34 34 30 38 Loop 34 34 30 38
Whorl 33 33 34 32 Whorl 33 33 34 32
Total 100 100 100 100 Total 100 100 100 100
Table 24 – Digital Live Scan Pattern Classification for Single Finger Images by Finger (Females)
Data From Females (Single Finger)
Pattern Class
R. Thumb
R. Index
R. Middle
R. Ring
R. Little
L. Thumb
L. Index
L. Middle
L. Ring
L. Little
Total
Arch 4 4 4 3 3 3 3 3 3 3 33
Loop 3 3 3 3 3 3 4 4 4 4 34
Whorl 3 3 3 4 4 4 3 3 3 3 33
Total 10 10 10 10 10 10 10 10 10 10 100
Table 25 – Digital Live Scan Pattern Classification for Single Finger Images by Finger (Males)
Data From Males (Single Finger)
Pattern Class
R. Thumb
R. Index
R. Middle
R. Ring
R. Little
L. Thumb
L. Index
L. Middle
L. Ring
L. Little
Total
Arch 4 4 4 3 3 3 3 3 3 3 33
Loop 3 3 3 3 3 3 4 4 4 4 34
Whorl 3 3 3 4 4 4 3 3 3 3 33
Total 10 10 10 10 10 10 10 10 10 10 100
Data demographics
The fingerprint images used to compile the datasets as described above were taken from several subjects. The
balancing of the samples used was based on the uniqueness of a single fingerprint and not individual subjects. As such,
multiple, yet distinct, fingerprint impressions were taken from some subjects (i.e., some subjects contributed more
than one finger).
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 48 OF 52
Table 26 - Gender Breakdown for Data
Subjects by Gender and Race
Males Females White Black Hispanic Asian
Ink Card Scan Dataset 72 17 38 47 3 1
Digital Live Scan Dataset 60 63 11 51 59 2
Combined Dataset (All) 132 80 49 98 62 3
Table 27 - Age Breakdown for Data
Subjects by Age
Under 18 18-24 25-29 30-34 35-39 40-44 45-49 50+
Ink Card Scan Dataset 5 54 18 3 1 4 3 1
Digital Live Scan Dataset 9 38 23 15 18 8 4 8
Combined Dataset (All) 14 92 41 18 19 12 7 9
Table 28 - Other Metadata: Height and Weight
Subjects by Height and Weight
<5’0”
5’0” – 5’5”
5’6” – 5’11”
6’0”+
<100 lbs
100 – 149 lbs
150 – 199 lbs
200 – 249 lbs
250+ lbs
Ink Card Scan Dataset 0 11 59 19 0 27 48 10 4
Digital Live Scan Dataset 4 45 63 11 1 45 56 17 4
Combined Dataset (All) 4 56 122 30 1 72 104 27 8
Table 29 - Other Metadata: Eye Color
Subjects by Eye Color
Brown Black Blue Green Hazel
Ink Card Scan Dataset 69 2 12 1 5
Digital Live Scan Dataset 111 1 2 6 3
Combined Dataset (All) 180 3 14 7 8
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 49 OF 52
Table 30 - Image Geometry Data
Impression Type Image Width (Pixels) Image Height (Pixels) Image Size (KB, Non-compressed)
Mean Median Min Max Mean Median Min Max Mean Median Min Max
Ink Card Scan Rolled 1016 1002 652 1718 1166 1165 643 2063 1170.8 1134.4 535.0 2568.6
Ink Card Scan Flat 602 592 444 843 785 801 497 1008 459.6 476.0 269.0 564.1
Ink Card Scan Slap 3192 3192 3045 3334 2009 2013 1744 2082 6264.4 6264.7 5339.3 6597.5
Digital Live Scan
Rolled
1600 1600 1600 1600 1500 1500 1500 1500 2343.8 2343.8 2343.8 2343.8
Digital Live Scan Flat 687 674 515 825 1057 993 672 1500 725.1 652.5 377.9 1171.9
Digital Live Scan Slap 3200 3200 3200 3200 2000 2000 2000 2000 6250.0 6250.0 6250.0 6250.0
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 50 OF 52
Appendix B. Equipment Used for Study
The equipment utilized for the processing of the image data with the algorithms described was comprised of a single
PC customized specifically for the purposes of imaging software development, research, and testing. The specifications
of this PC are as follows:
Model: Dell Precision T7500
CPU: 2x Intel Xeon W5580 @ 3.20 GHz
Memory: 12.0 GB DDR3 (Registered ECC)
Storage: 2x OCZ Vertex 2 240GB SSD (RAID 1), 4x Western Digital 2TB HDD (RAID 0+1)
Operating System: Microsoft Windows 7 64-bit Ultimate Edition
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 51 OF 52
Appendix C. Examination of Entropy
Lossy compression may balance compressed file size against the discard of image information. That is, a lossy CODEC
may achieve a desired compression ratio by varying the amount of information discarded, or achieve a specified level of
fidelity to the original image by altering the size of the output compressed file, i.e., a lower compression ratio. Various
CODECs provide varying degrees of user control over this balance, but in every case, the complexity of image content
will set the stage for various optimization strategies to be executed by the compression algorithm in balancing
information loss against compression ratio.
By contrast, lossless image compression, by definition, must preserve all information contained in the non-compressed
original. Loss of information is not available to the lossless CODEC as part of its optimization scheme. Accordingly, the
only option available to the lossless algorithm is to increase the compressed file size (i.e., reduce the compression ratio)
in proportion to the complexity of the input non-compressed original.
In the present study, it was observed that the BMP-RLE CODEC performed reasonably well with the Digital Live Scan
images, but poorly with the Ink Card Scan images, actually resulting in a compressed file size larger than that of the
non-compressed original. Suspecting that the RLE effectiveness to be related directly to image entropy differences, we
measured the entropy of both Ink Card Scan and Digital Live Scan for comparison.
Run-length encoding (RLE) is a very simple form of data compression in which runs of data (sequences in which the
same data value occurs in many consecutive data elements) are stored as a single data value and count, rather than as
the original run. Hence, images containing large regions of homogeneous gray level may be encoded very efficiently
using RLE. By contrast, highly textured images containing few runs are less efficiently encoded via RLE.
Entropy (E) provides a measure of the average gray level variability of an image. It is defined as:
2
1
( )log ( )
n
I i i
i
E p x p x
(C.1)
where n = number of gray levels in the image, i.e. 256; xi = the value of the i
th gray level; p(xi) = the probability of
occurrence of the i
th gray level in the image. Thus, for a single channel (8 bit) image having p(xi) = 1/256 for every i, the
maximum entropy is 8, or 8 bits. Accordingly, an image displaying a relatively flat (uniform) gray level histogram should
yield an entropy very near the maximum value of 8 bits. An image having less uniform distribution of gray levels, such as
an image displaying large areas homogeneous in gray level, would be expected to yield a lower entropy measure.
Procedure
Entropy was measured for each of the non-compressed source images used in the present study. This included 1000
ppi digital Scans acquired from: a standard inked 10-print fingerprint cards, and a Digital Live Scan device employing the
principal of Frustrated Total Internal Reflection (FTIR).
In addition to entropy as defined in equation (1), we measured the ratio of entropy in the image area identified by a
segmentation procedure as occupied by the background to that of the fingerprint. Hence, we have the Entropy Ratio
defined as simply:
ingerprint
background
Ratio
f
E
E
E
(C.2)
NISTIR 7779 - An Exploration of the Operational Ramifications of Lossless Compression of 1000 ppi Fingerprint Imagery
PAGE 52 OF 52
Results
Table 31 exhibits mean values of various entropy measures for Ink Card Scan data and those acquired using an FTIR
Digital Live Scan device. The table resolves the mean values for each of the fingerprint impression types, rolled, flat, and
4-finger slap. Also included are measures of the proportion (percentage) of the image area identified via the
segmentation procedure as containing the fingerprint.
Table 31 – Mean measurements of entropy and fingerprint region of images
CODEC
Entropy Image Entropy
Fingerprint
Entropy
Background
Entropy Ratio Prop Area
Fingerprint
Rolled Ink Card Scan 7.6 7.6 6.4 0.848 85.3
Digital Live Scan 3.2 6.5 0.1 0.019 41.0
Flat Ink Card Scan 7.4 7.4 6.0 0.811 63.7
Digital Live Scan 5.7 6.9 1.6 0.227 66.3
Slap Ink Card Scan 7.3 7.6 6.5 0.855 39.7
Digital Live Scan 4.4 7.2 0.8 0.110 43.7
Entropy of the images tends to be higher for Ink Card Scan images in contrast to that of Digital Live Scan images.
Entropy of fingerprint regions tends to be slightly lower for Digital Live Scan, but the main contrast is between entropy
of the background regions, where that of Digital Live Scan is substantially lower than that of the Ink Card Scan.
Examination of the measures of the proportion of image area occupied by the fingerprint shows the greatest difference
with the rolled impressions. This impression type, thus, is most likely to contain large areas of low-entropy background
most efficiently compressed using the RLE algorithm. Such regions are more easily compressed by other algorithms as
well yielding the highest effective compression ratios, i.e., smaller compressed file sizes.
