Effective Image Compression using Evolved Wavelets

Uli Grasemann
Department of Computer Sciences The University of Texas at Austin Austin, TX 78712

Risto Miikkulainen
Department of Computer Sciences The University of Texas at Austin Austin, TX 78712

[email protected]

[email protected]

ABSTRACT
Wavelet-based image coders like the JPEG2000 standard are the state of the art in image compression. Unlike traditional image coders, however, their performance depends to a large degree on the choice of a good wavelet. Most wavelet-based image coders use standard wavelets that are known to perform well on photographic images. However, these wavelets do not perform as well on other common image classes, like scanned documents or ngerprints. In this paper, a method based on the coevolutionary genetic algorithm introduced in [11] is used to evolve specialized wavelets for ngerprint images. These wavelets are compared to the hand-designed wavelet currently used by the FBI to compress ngerprints. The results show that the evolved wavelets consistently outperform the hand-designed wavelet. Using evolution to adapt wavelets to classes of images can therefore signicantly increase the quality of compressed images.

1. INTRODUCTION
Image compression is one of the most important and successful applications of the wavelet transform. Mature waveletbased image coders like the JPEG2000 standard [15] are available, gaining in popularity, and easily outperform traditional coders based on the discrete cosine transform (DCT) like JPEG [25]. Unlike in DCT-based image compression, however, the performance of a wavelet-based image coder depends to a large degree on the choice of the wavelet. This problem is usually handled by using standard wavelets that are not specially adapted to a given image, but that are known to perform well on photographic images. However, many common classes of images do not have the same statistical properties as photographic images, such as ngerprints, medical images, scanned documents, and satellite images. The standard wavelets used in image coders often do not match such images, resulting in decreased compression or image quality. Moreover, non-photographic images are often stored in large databases of similar images, making it worthwile to nd a specially adapted wavelet for them. As Chris Brislawn, one of the architects of WSQ [13], the FBIs standard for ngerprint compression, states [2]: Choosing wavelets for image coding applications is still a somewhat inexact science, depending on a lot of trial and error. There are a few standard wavelet families [...] that seem to work well for image coding, although that is not a task for which they were specically designed. In the future we hope to be able to design wavelets (or wavelet-like lter banks) that are optimized for a specic application, like ngerprints. Until then well probably stick with proven performers [...]. In this paper, a coevolutionary genetic algorithm based on Enforced Sub-Populations [11, 9] and a mathematical technique called Lifting is used to nd wavelets that are specially adapted to a particular class of images. The approach is tested in the ngerprint compression domain, which provides a systematic comparison to other current approaches. The wavelets found by the GA are tested in a state-of-the-art image coder, and compared with standard wavelets, including the winner of a competition held by the FBI to nd the best wavelet for ngerprint compression [13]. The evolved wavelets turn out consistently better, demonstrating that evolutionary discovery can outperform human design in an important task. The remainder of this paper is structured as follows. The

Categories and Subject Descriptors

G.1.2 [Numerical Analysis]: ApproximationWavelets and Fractals; I.4.2 [Computing Methodologies]: Image Processing and Computer VisionCompression (Coding); I.2.6 [Computing Methodologies]: Articial Intelligence Learning; I.2.8 [Computing Methodologies]: Articial IntelligenceProblem Solving, Control Methods, and Search; G.1.6 [Numerical Analysis]: Global Optimization

General Terms
Algorithms, Experimentation, Performance

Keywords
Wavelets, Image Compression, Lifting, Genetic Algorithms, Coevolution

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for prot or commercial advantage and that copies bear this notice and the full citation on the rst page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specic permission and/or a fee. GECCO05, June 2529, 2005, Washington, DC, USA. Copyright 2005 ACM 1-59593-010-8/05/0006 ...$5.00.

0.8

1.6

0.6

1.2

0.4

0.8

0.2

0.4

-0.2

-0.4

-0.4

-0.8

-3

-2

-1

-3

-2

-1

Figure 1: An example wavelet function. This wavelet is the Antonini 9/7 wavelet [1], also known as the FBI wavelet, because the FBI uses it to compress ngerprints [13]. next section gives a brief tour of wavelet-based image compression. Section 4 describes the algorithm and the evaluation function used. Section 5 reports experimental results, and section 6 discusses future directions for the wavelet evolution approach.

Figure 2: A translated and dilated version of the Antonini wavelet in gure 1. The Discrete Wavelet Transform (DWT) expresses the input data as a weighted sum of such translates and dilates. we have a discrete wavelet transform or DWT, essentially leading to a nite number of coecients. In order to compute the DWT of a function f , we need to nd one wavelet coecient j,i for each j,i , such that f=
j,i

j,i j,i .

(3)

2.

BACKGROUND

2.1 Wavelets
Wavelets are a mathematical tool for representing and approximating functions hierarchically. At the heart of wavelet theory, there is a single function , called the mother wavelet. Any function can be represented by superimposing translated and dilated versions of , denoted by j,i , where i and j are the translation and dilation parameter. We are focusing on the discrete case where i and j only take on integer values. The j,i can be computed from the mother wavelet as j,i (x) = 2 2 (2j x i).
j

where the bar denotes the complex conjugate. Otherwise, a dual wavelet is necessary such that and together are biorthogonal, which basically means that the transform must be invertible. We can then use for determining the wavelet coecients (equation 4), and the original wavelet for the inverse DWT (equation 3). Note that an orthogonal wavelet is just a special case of a biorthogonal one where = .

2.2 Filters and the Fast Wavelet Transform

Computing a wavelet transform in the way just described is expensive and cumbersome. However, an algorithm called the Fast Wavelet Transform or FWT allows computing the wavelet coecients by recursively applying a pair of digital lters to the data, much like the Fast Fourier Transform reduces a discrete fourier transform to computing a few nite sums. A digital lter can be dened by giving a sequence of real numbers called lter coecients. It is applied by convolution with an input sequence. A lter is said to have nite impulse response (FIR), if its coecients are non-zero only on a nite range. A FIR lter can be represented by a nite number of coecients and the index of the leftmost non-zero coecient. The lter pair used in the FWT uniquely determines the mother wavelet and also (in the biorthogonal case) the dual wavelet . In order to dene a valid wavelet transform, a lter pair must be complementary, which is the same as saying that the associated wavelet must be biorthogonal.

(1)

Figure 1 shows an example wavelet, and gure 2 shows a translated and dilated version of that wavelet. All the translates of for a specic dilation j span a function space Wj :

Wj = span{ j,i | i

}.

(2)

The Wj are called wavelet spaces or detail spaces, because each of them adds a level of detail to the wavelet representation of a function. All of the detail spaces combined form a basis in which any function can be expressed. The process of decomposing a function into wavelet coecients (a scaling factor for each of the j,i ) is called wavelet transform. If the parameters i and j take on dicrete values,

Both wavelet theory and wavelet-based image compression are complex and evolving subjects. This section gives a brief high-level overview of these topics. For more details on classical wavelet theory, see [14]. In addition, [6] contains an introduction to lifting, and [7] covers the basics of wavelet-based image compression.

If a wavelet basis (i.e. the set of all j,i ) is orthogonal, then the j,i are given by

j,i =

f, j,i

f (x)j,i (x)dx,

(4)

Ensuring that a lter pair is complementary and that the individual lters are nite are the basic contraints in wavelet design.

2.3 Lifting
The Lifting scheme, introduced by Sweldens [22] in 1996, oers an eective way to construct complementary lter pairs. A nite lter, called a lifting step, is used to generate a new lter pair from an existing pair. Multiple lifting steps can be applied consecutively. In [6], Sweldens and Daubechies proved two important properties of lifting: Lifting preserves biorthogonality, i.e. if the original lter pair is complementary, then so is the new pair, no matter what lifting step is applied. Any wavelet with nite lters can be expressed as a sequence of lifting steps. Starting with the trivial wavelet transform (called the Lazy Wavelet), all possible wavelets can be reached by applying a nite number of nite-length lifting steps. These two properties make lifting a powerful tool for constructing new wavelets: Starting from a known complementary lter pair, other complementary pairs that are better adapted to the task at hand can be generated by applying lifting steps.

All three steps of an image coder have an impact on the image quality achieved for a given compression ratio. In particular, a better transform results in better compression performance.

3. RELATED WORK
Adaptive wavelet bases have been an active research area since the early 1990s. Traditional approaches are based on dictionary methods, where a basis is selected from an overcomplete set of predened functions called atoms. Examples of such methods are the best basis algorithm [5] and wavelet packets [26]. In [16] and [18], evolutionary algorithms are used for adaptive dictionary methods. Dictionary methods are in a sense orthogonal to the approach used in this paper, and could be easily combined with it. Several stochastic optimization techniques have been applied to the design of wavelets. Monro and Sherlock [20] use simulated annealing to nd wavelets with balanced uncertainty in space and frequency. Hill et al. [12] used a genetic algorithm to optimize the parameters of a windowed trigonometric function that can be used in a continuous wavelet transform. The lifting technique has provided new ways to adapt wavelets to the data being transformed. The common approach is to choose locally between a class of lifting steps while transforming an image. For example Claypoole et al. [3] used this approach to adapt wavelets to a given signal by optimizing data-based prediction error criteria. Genetic algorithms, combined with the lifting technique, oer a way to nd a single wavelet specically adapted to a class of images based on real-world compression performance. A rst version of such an algorithm was introduced in [11], and applied to a simple signal compression problem. In this paper, the method is shown to be eective in the more challenging task of compressing ngerprint images.

2.4 Wavelet-based Image Compression

All modern image coders are transform coders, i.e. they have the structure shown in gure 3. Transform coders rst apply an invertible transform to the image data in order to decorrelate it. Examples of such transforms are the discrete cosine transform (the basis for JPEG compression), and the discrete wavelet transform, the basis for JPEG2000 and other wavelet coders. The performance of a transform coder depends largely on how well the transform decorrelates the signal. A well decorrelated signal consists mainly of coecients close to zero.

4. EVOLVING WAVELETS
In section 2.3, two useful properties of lifting were mentioned: Lifting preserves biorthogonality, and

Trans formation

Quanti zation

Entropy Coding

0100110...

any wavelet can be expressed as a sequence of lifting steps. These two properties make sequences of lifting steps an effective representation for wavelets in a genetic algorithm, because (1) any random sequence of lifting steps will encode a valid (i.e. biorthogonal) wavelet, and (2) any wavelet can be represented using the genetic code. In this section, a coevolutionary genetic algorithm that evolves wavelets encoded as lifting steps will be described.

RD Control

Figure 3: The structure of a transform coder. The signal is rst decorrelated using an invertible transform, then quantized and entropy coded. The ratedistortion (RD) unit controls the quantization to minimize the distortion within the available bit rate. The performance of a transform coder depends on how well the transform decorrelates the image data. After the transform step, the coecients are quantized, i.e. expressed using symbols from a nite alphabet, and entropy coded, using as little space or bandwidth as possible. The rate-distortion (RD) unit controls the quantization in order to achieve minimal distortion within the available bit rate.

Algorithm
The coevolutionary GA used is closely related to the Enforced Sub-Populations (ESP) neuroevolution algorithm introduced by Gomez and Miikkulainen [9]. ESP evolves a number of populations of individual neurons in parallel. In the evaluation phase, ESP repeatedly selects one neuron from each sub-population to form candidate networks. The tness of a particular neuron is the average tness of all networks in which it participated. This concept can be easily applied to wavelet evolution: Several populations of lifting steps are evolved in parallel,

Mutation
Wavelet-ESP input: N, the number of sub-populations L, the lengths of the lifting lters M, the size of each sub-population P, the mutation rate 1. Initialize Create M lters of length L for each of the N sub-populations, and randomize them. 2. Evaluate Select N lifting steps, one from each sub-population, and evaluate the resulting wavelet. Add the tness to the cumulative tness of all participating steps. Repeat until each step has been evaluated 10 times on average. 3. Recombine Rank the lifting steps in each sub-population by their average tness. Each step in the top quartile is recombined with a higher-ranking step. The ospring is mutated with probability P and replaces the lowest-ranking half of each sub-population. 4. Repeat Repeat the Evaluate-Recombine cycle for a xed number of generations. Figure 4: The ESP algorithm applied to wavelets. Several populations of lifting steps are evolved in parallel, and are combined in the evaluation phase to form wavelets. and are randomly combined to form wavelets, which are then evaluated. No migration or crossover occurs between sub-populations. Figure 4 describes the algorithm in detail. A chromosome is mutated by adding low-variance gaussian noise to a random lter coecient and/or adding 1 to the integer representing the leftmost index.

Fitness Evaluation
In image compression, the ideal measure of tness would be the performance in an actual transform coder as described in section 2.4. However, there are two problems with this approach. First, evaluating a wavelet using a transform coder is prohibitively expensive. Second, in order to make a fair comparison between two wavelets, either the available number of bits needs to be xed and the resulting distortion used as a tness measure, or vice versa. Both options are inexact and expensive for actual transform coders.

Evaluation Function input: D, the input data W, a candidate wavelet R, the compression ratio return: The tness of W.

1. Transform Transform D using the wavelet W. 2. Compress Sort the resulting wavelet coecients. Keep only the largest R |D|. Set the rest to zero. 3. Reconstruct Perform an inverse transform using W and the altered wavelet coecients. 4. Measure the Image Quality Measure the resulting image quality (peak signal to noise ratio), and return it. Figure 5: The evaluation function is an idealized version of a transform coder: Instead of quantizing and entropy-coding the wavelet coecients, it uses only part of the coecients for reconstruction and sets the rest to zero. Figure 5 shows a denition of the evaluation function. It is an idealized version of a transform coder: Instead of quantizing and entropy-coding the wavelet coecients, it uses only a certain percentage of the coecients for reconstruction and sets the rest to zero. This approach is much less expensive and allows choosing the compression ratio exactly, which means that the resulting distortion can be used directly as a tness measure. Villasenor et al. [24] have used a similar but even simpler method to evaluate wavelets with good results. Figure 6 compares the performance predicted by the evaluation function to the actual performance in a transform coder, using wavelets tested during the experiments reported in the next section. The gure shows that the prediction

Representation
A lifting step is represented as a xed-length sequence of oating point numbers for the lter coecients, and a single integer for the leftmost index of the lter. Using a xed number of xed-length steps limits the number of wavelets that can be represented. However, it also limits the length of the wavelet lters, which is a desirable eect. Also, most wavelets used in practice can be factored into a small number of short lifting steps [6], so this limitation is unlikely to interfere with nding good solutions.

Initialization
Each chromosome is initialized by setting the values of the coecients to random values from a gaussian distribution with mean 0 and variance 0.5, and setting the leftmost index of each lter to a random integer between -2 and 2. These settings reect the values commonly found in lifting steps.

Crossover
The crossover operator performs simple one-point crossover on the coecients. The integers representing the leftmost indices of the parent lters are randomly assigned to the children.

35

30
Actual Performance (PSNR)

25

20

15

10 10 15 20 25 Predicted Performance (PSNR) 30 35

Figure 6: The relation between the performance predicted by the evaluation function (gure 5), and real-world performance in a transform coder. Each point represents a wavelet encountered in the experiments reported in section 5. of the real-world performance is accurate, although in a few cases, the evaluation function underestimates the actual performance.

Figure 7: The FBI digitizes and compresses between 30,000 and 50,000 new ngerprint cards every day. The images shown are part of the data set used in the FVC2000 ngerprint verication competition [19], and were also used in the reported experiments.

Parameters
Preliminary experiments were conducted to determine the best parameter settings. The algorithm turned out to be very robust; similar results were obtained for a wide range of parameters. The following parameters were used for the reported results: The population size was 150 for each sub-population, which means that 1500 evaluations took place in each generation. The algorithm evolved 7 sub-populations in parallel, each of which contained lifting steps of length 4. The mutation rate was set to 0.4. The compression ratio was 16:1 both for tness evaluation and the evaluation on the test image. The evolution ran for 500 generations each time.

5.

EXPERIMENTS

In this section, the wavelet evolution method described in the last section is applied to the compression of ngerprint images like the ones shown in gure 7.

5.1 Task and data

Fingerprint compression is both a popular research problem in image compression and an important application in its own right. The FBI alone has over 200 million ngerprint cards on le, occupying an acre of ling cabinets[2]. Recently, they have begun to store ngerprints electronically. Between 30,000 and 50,000 new cards are digitized and compressed every day, using a wavelet-based image coder and the so called FBI wavelet, considered to be the best known wavelet for ngerprint compression [13]. The FBI wavelet (shown in gure 1) therefore provides a challenging benchmark with which to compare the evolved wavelets. For the experiments reported in this section, the rst set of ngerprints from the FVC2000 nger print verication competition [19] was used. The data set contains 80 blackand-white images acquired electronically using an optical sensor (Secure Desktop Scanner by KeyTronic). The size of each image is 300 by 300 pixels, at 500 dpi resolution.

Quality Analysis
After each generation, the best wavelet found so far was used to compress the test image. The image coder was the implementation of the Set Partitioning in Hierarchical Trees (SPIHT) algorithm [21] in J. E. Fowlers QccPack, an opensource library of state-of-the-art data compression routines [8]. The performance on the test set can therefore be regarded as an accurate measure of real-world performance. The following analysis relies on error images and peak signal-to-noise ratio (PSNR) as measures of image quality. The PSNR is a simple logarithmic measure for the difference between two images. It is commonly used in image compression as a quantitative measure for the compression error introduced by a compression algorithm. A PSNR of 30 decibel is commonly regarded as a reasonable lower quality limit, and a PSNR above 50dB is regarded as visually perfect, i.e. the compressed image is visually indistinguishable from the original. A dierence of 0.40.5 dB between two algorithms is usually visible. Error images are obtained by subtracting the compressed image pixel-by-pixel from the original. Error images are necessary because subtle dierences in quality between compressed images are very hard to judge visually. In print, even substantial quality improvements are unlikely to be clearly visible. Separating the error from the image data makes comparisons much easier.

5.2 Methodology
Cross-Validation
The algorithm was evaluated using leave-one-out cross validation on the 80 available images, i.e. each of the images was used once as a test image, and 79 times as part of the training set. Each of the 80 runs took approximately 45 minutes on a 3GHz Xeon processor.

Generation 1

Generation 10

Generation 20

Generation 50

Figure 8: The progress of evolution during a typical run is shown at generations 1, 10, 20 and 50. The top row shows the winner wavelets, and the bottom rows shows the resulting compressed test image at 16:1. The rst generation produced a more or less random wavelet that performs poorly. Over the next generations, both image quality and the smoothness of the wavelets increase sharply. The performance keeps increasing after generation 50 (gure 9), although the dierences are less obvious.

5.3 Results
Figure 8 illustrates the progress of evolution during a typical run of the algorithm. The winner wavelets and the resulting compressed test images are shown after generations 1, 10, 20, and 50. The winner of the rst generation is highly discontinuous and performs poorly. As the wavelets adapt to the structure of the data, they become smoother, and the image quality increases sharply. The performance keeps improving well beyond generation 50, although the dierences in image quality are less obvious. Figure 9a shows the learning curve of the algorithm, i.e. the performance on the test images, averaged over all 80 runs of the algorithm. The horizontal lines show the average performance of the FBI wavelet and a baseline JPEG coder. The curve shows that the evolved wavelets achieve an average improvement of 0.75dB over the FBI wavelet. The FBI wavelet in turn outperforms JPEG by approximately 2.5dB. Alternatively, a 0.75dB quality improvement would translate into a 15-20% decrease in space requirements for the same image quality. Figure 9b compares the evolved wavelets to the FBI wavelet directly. The dotted line shows the lower limit of the 95% condence interval. At 30 generations, the evolved wavelets perform the same on average as the FBI wavelet. By generation 40, there is a 95% probability of nding a better wavelet than the FBI wavelet. After 500 generations, the evolved wavelets are 95% certain to outperform the FBI wavelet by more than 0.45dB. At this point, the evolved wavelet is signicantly better with p < 104 , according to a paired t-test. Figure 10 gives a visual impression of the increase in image quality achieved by the evolved wavelets. The top row shows a detail of the leftmost ngerprint in gure 7, compressed at 16:1 using the wavelet found by the GA, the FBI wavelet,

and a baseline JPEG coder. Both wavelets perform much better than the JPEG coder: The blocking artifacts that often accompany JPEG-encoded images are clearly visible. The dierence between the two wavelets is clearer in the bottom row, which shows only the error introduced by each compression method. The error image of the evolved wavelet is dimmer than that of the FBI wavelet, indicating a more accurate reconstruction of the original image. The constrast of the error images was increased uniformly to enhance the visibility of the non-zero dierences.

6. DISCUSSION AND FUTURE WORK

The results presented in the previous section show that evolving wavelets for classes of images can considerably improve the performance of an image coder. The evolved wavelets substantially outperformed the hand-optimized FBI wavelet in every single run of the algorithm. Many other image classes, including medical images, structural drawings, digitized documents, and satellite images have the characteristics that would make it useful to evolve specialized wavelets for them: They are often stored in large databases of similar images, they do not have the same statistical structure as photographs, and they have regularities of their own that standard wavelets cannot fully exploit. The results presented in this paper suggest that similar improvements should be possible in these cases. Other applications of the same algorithm, like lossless image compression, compression of volumetric data or wavelet-based multi-grid solvers for partial dierential equations, are also possible in the future. The algorithm itself could also be extended to evolve more powerful classes of wavelets. The design of non-separable and nonlinear wavelet transforms has received much attention in the literature recently (e.g. [10, 4, 23]). The algo-

33 Evolved wavlet + SPIHT coder FBI wavelet + SPIHT coder JPEG 32

1.2 Evolved wavelet vs. FBI wavelet p < 0.05 1

Image Quality (PSNR)

31

Quality Improvement (PSNR)

0.8

30

0.6

29

0.4

28

0.2

27 0 50 100 150 200 250 Generations 300 350 400 450 500

0 0 50 100 150 200 250 Generations 300 350 400 450 500

(a) Learning curve.

(b) Evolved vs. FBI wavelet.

Figure 9: Learning curve of the wavelet evolution algorithm. In each generation, the best wavelet found was used in a state-of-the-art image coder. Plot (a) shows the resulting image quality on the test image, averaged over 80 runs. The horizontal lines show the performance of the FBI wavelet and a baseline JPEG coder. Plot (b) shows the average quality improvement over the FBI wavelet. The dotted line is the lower limit of the 95% condence interval. After 500 generations, the evolved wavelets are 95% certain to outperform the FBI wavelet by at least 0.45dB, a dierence usually visible to the naked eye. rithm used in this paper could be adapted to evolve nonseparable and nonlinear wavelets without major changes. [5] R. Coifman and V. Wickerhauser. Entropy-based algorithms for best basis selection. IEEE Transactions on Information Theory, 38:713718, 1992. [6] I. Daubechies and W. Sweldens. Factoring wavelet transforms into lifting steps. Journal of Fourier Analysis and Applications, 4(3):245267, 1998. [7] G. Davis and A. Nosratinia. Wavelet-based image coding: An overview. Applied and Computational Control, Signals and Circuits, 1(1), 1998. [8] J. E. Fowler. Qccpack: An open-source software library for quantization, compression, and coding. In e. A. G. Tescher, editor, Applications of Digital Image Processing XIII, volume Proc. SPIE 4115, pages 294301, August 2000. [9] F. Gomez and R. Miikkulainen. Solving non-markovian control tasks with neuroevolution. In Proceedings of the International Joint Conference on Articial Intelligence, pages 13561361, San Francisco, CA, 1999. [10] A. Gouze, M. Antonini, M. Barlaud, and B. Macq. Design of signal-adapted multidimensional lifting scheme for lossy coding. IEEE Transactions on Image Processing, 13(12):1589 1603, December 2004. [11] U. Grasemann and R. Miikkulainen. Evolving wavelets using a coevolutionary genetic algorithm and lifting. In Proceedings of the Genetic and Evolutionary Computation Conference, volume II, pages 969980, New York, NY, 2004. Springer. [12] Y. Hill, S. OKeefe, and D. Thiel. An investigation of wavelet design using genetic algorithms. In Microelectronic Engeneering Research Conference, 2001. [13] T. Hopper, C. M. Brislawn, and J. N. Bradley. WSQ Gray-scale Fingerprint Image Compression Specication. Federal Bureau of Investigation, February 1993. Document No. IAFIS-IC-0110.

7.

CONCLUSIONS

In this paper, a coevolutionary GA was used to evolve specialized wavelets for image compression, using ngerprint images as a test domain. The evolved wavelets consistently outperform the wavelet used by the FBI in this task. These results show that evolving wavelets adapted to specic image classes can signicantly increase the compression performance of an image coder. They also demonstrate that evolutionary discovery can outperform signicant human design eort in an important task.

8.

ACKNOWLEDGEMENTS

This research was supported in part by the National Science Foundation under grant IIS-0083776 and by the Texas Higher Education Coordinating Board under grant ARP003658-476-2001.

9.

REFERENCES

[1] M. Antonini, M. Barlaud, P. Mathieu, and I. Daubechies. Image coding using wavelet transform. IEEE Transactions on Image Processing, 1992. [2] C. M. Brislawn. The fbi ngerprint image compression standard. http://www.c3.lanl.gov/brislawn/FBI/FBI.html. [3] R. Claypoole, R. Braniuk, and R. Nowak. Adaptive wavelet transforms via lifting. Transactions of the International Conference on Acoustics, Speech and Signal Processing, pages 15131516, May 1998. [4] R. Claypoole, G. Davis, W. Sweldens, and R. Baraniuk. Nonlinear wavelet transforms for image coding via lifting. IEEE Transactions on Image Processing, 12(12):1449 1459, December 2003.

(a) Original

(b) Evolved

(c) FBI Wavelet

(d) JPEG

Figure 10: (a) A detail of the left ngerprint in gure 7, compressed at 16:1 with an evolved wavelet (b), the FBI wavelet (c), and a baseline JPEG coder (d). The bottom row shows the error introduced by each compression method. The error image of the evolved wavelet is dimmer than that of the FBI wavelet, reecting the lower overall distortion. [14] B. Jawerth and W. Sweldens. An overview of wavelet based multiresolution analyses. SIAM Rev., 36(3):377412, 1994. [15] I. JTC1/SC29/WG1. JPEG 2000 - lossless and lossy compression of continuous-tone and bi-level still images. Part I: Minimum decoder. Final committee draft, version 1.0., March 2000. [16] M. M. Lankhorst and M. D. van der Laan. Wavelet-based signal approximation with genetic algorithms. In Evolutionary Programming, pages 237255, 1995. [17] R. L. Lippke. The disenfranchisement of felons. Law and Philosophy, 20:553 580, Nov 2001. [18] C. Liu and H. Wechsler. Face recognition using evolutionary pursuit. In Proceedings of the Fifth European Conference on Computer Vision, Freiburg, Germany, 1998. [19] D. Maltoni, D. Maio, A. Jain, and S. Prabhakar. Handbook of Fingerprint Recognition. Springer, New York, June 2003. [20] D. Monro and B. Sherlock. Space-frequency balance in biorthogonal wavelets. Transactions of the IEEE Int. Conf. on Image Processing, 1:624627, 1997. [21] A. Said and W. Pearlman. A new fast and ecient image codec based on set partitioning in hierarchical trees. IEEE Transactions on Circuits and Systems for Video Technology, 6:243250, June 1996. [22] W. Sweldens. The lifting scheme: A custom-design construction of biorthogonal wavelets. Journal of Applied and Computational Harmonic Analysis, 3(2):186200, 1996. [23] W. Sweldens. The lifting scheme: A construction of second-generation wavelets. SIAM J. Math. Anal., 29(2):511546, 1997. [24] J. Villasenor, B. Belzer, and J. Lia. Wavelet lter evaluation for image compression. IEEE Transactions on Image Processing, 2:10531060, August 1995. [25] G. Wallace. The jpeg still picture compression standard. IEEE TCE, 38, 1992. [26] M. Wickerhauser. Adapted Wavelet Analysis from Theory to Software. A. K. Peters, Wellesley, MA, 1994.