Book review: Handbook of Iris Recognition
James L. Wayman · IET Biometrics · 2014
Fifteen years ago, it was hard to find more than a handful of scientists with interest and experience in iris recognition. This year, as part of their Advances in Computer Vision and Pattern Recognition series, Springer has published Handbook of Iris Recognition, which contains 19 contributed chapters from 42 different researchers, plus a forward by the 'father' (or perhaps after more than 20 years devoted to advancing the art and science, now the 'grandfather') of automated approaches to iris recognition, Prof. John Daugman. It is impressive how this human recognition technology has matured scientifically and found a commercial market in such a relatively short time, due in large measure to the effective research, development and advocacy of Prof. Daugman. The scientific development of the concept of recognising persons by their iris patterns dates to at least Alfonse Bertillon's work in the late nineteenth century. Bertillon's 1886 treatise on 'The Color of the Iris' was motivated by the potential use of iris colour and texture as a means of differentiating individuals. In 1987, ophthalmologists Leonard Flom and Alan Safir obtained a broad patent (US Patent 4,641,349) on a method and apparatus for automated iris recognition, but proposed simple pixel-by-pixel correlation comparisons of iris images. It was John Daugman who, working with Flom and Safir, proposed and patented (US Patent 5,291,560; 1994) a truly ingenious method for comparing iris patterns making iris recognition systems a practical reality. Both the Flom-Safir and Daugman patents were licensed in the 1990s to a series of companies, some of which aggressively sought to prevent commercial competition and discourage third-party scientific research not consistent with industry claims of zero observed errors. This position hindered both the scientific development and the broad adoption of iris recognition technologies, particularly by government agencies. With the expiration of the Flom and Safir patent in 2008, iris recognition took off dramatically. As the Handbook Forward, Chapter 1, 'Introduction to the Handbook of Iris Recognition', and Chapter 2, 'A Survey of Iris Biometric Research: 2008–2010' all point out, more scientific papers were published in the 2008–2010 time period than in the entire 15 years prior. Scientific activity has increased to the point that we now can have this comprehensive Handbook of Iris Recognition from an international collection of professional researchers not commercially linked to the technologies. The 19 chapters span a wide variety of concerns, from a survey of 221 iris recognition papers published from 2008 to 2010 (Chapter 2) through visible wavelength and cross wavelength iris recognition (Chapters 8 and 9) and spoofing by the reverse engineering of iris codes (Chapter 18) to a concluding article (Chapter 19) on the optical requirements for iris imaging. The Forward is written by Prof. Daugman, himself, and recalls the 'contentious' debates in the 1990s regarding 'the claim that the iris recognition algorithm had extraordinary resistance to false matches'. In my recollection, the debates, although indeed contentious, were more nuanced than simply whether or not iris recognition had low false match rates. The issue was whether the claims of false match rates of 1 in 13.5 billion (at HD = 0.25 in the 1994 patent) could be supported by comparisons of 592 irises from 323 persons. In contention was the appropriateness of unbridled extrapolation of the tails of a presumed binomial distribution with parameters that seemed to vary greatly from publication to publication (173 ≤ N ≤ 249). The Forward in the Handbook now proposes a false match rate of 1 in 1 million (at HD = 0.32), which is almost an order of magnitude better than suggested in the original 1994 patent (1 in 131,000 at HD = 0.321) and early publications (1 in 151,000 at HD = 0.32 in 1993), but considerably worse than offered in publications of the early 2000s (1 in 28M at HD = 0.32in 2003; 1 in 18M in 2004; not reported at HD = 0.32 in 2006, but greater than 1 in 1M; 1 in 26M at HD = 0.32, undated). With the sole exception of Chapter 8, 'Visible Wavelength Iris Recognition', the chapters of the Handbook avoid any speculative statements regarding false match rates lower than 1 in 1M, staying well within the bounds of the current empirical data. But there can be no question that the original approach of Prof. Daugman, that of extracting binary features from local phase information, was both effective and ingenious and, as Chapter 16, 'Introduction to IrisCode Theory' points out, is still 'the most successful iris recognition method'. It is quite remarkable that someone (in this case, Prof. Daugman) could have gotten it this right the first time. Chapter 2 summarises some of the non-Daugman-based academic algorithms, but the remaining chapters stay close to the old adage, 'If it works, don't fix it', keeping amazingly faithful to the original 1994 concept. Chapter 16 develops the theory of Daugman Iris Codes and shows why these codes are distinctive and efficient. Chapter 6, 'Iris Recognition with Taylor Expansion Features', proposes a variant to Gabor filters, but maintains the concept of local phase measurement and binary encoding, both of which are unique in biometrics to iris recognition (excepting a very few recent academic papers on the extraction of binary features from other modalities). Chapter 17, 'Application of Correlation Filters for Iris Recognition' by authors from Carnegie Mellon University, is a welcome break from the Daugman orthodoxy, developing floating point correlation metrics from the global two-dimensional Fourier-based transforms of the iris image. The editors promise a 'more sober appraisal of the field than exists in the market literature of the industry'. This promise is kept, particularly in Chapter 11, 'Template Aging in Iris Biometrics', that reaches conclusions running counter to the historic marketing literature (and, in fact, other chapters in this Handbook) claiming that iris texture is completely stable over time and therefore requires 'only a single enrolment in a lifetime'. In a well-controlled, 4-year study of 23 college students, faculty and staff, the University of Notre Dame authors demonstrate that comparison scores do degrade with increasing time between enrolment and comparison. This work follows widely discussed and circulated conference and journal articles on this subject by the same authors, albeit based on an even smaller database of images. This new study reaches the same conclusion that iris texture does not remain completely unchanged over time, as evidenced with degrading scores for mated comparisons with increasing time lapse between enrolment and test. But for the marketing claims, this would have been a predictable finding as we know through common experience that all human characteristics are subject to aging. Nothing in the chapter, however, indicates that iris recognition cannot be performed successfully over multi-year intervals. We are left wondering about what effects might be observed over longer time spans and with data subjects not chosen from a University population. I have expectantly awaited the publication of Chapter 5, 'Quality and Demographic Investigation of ICE 2006' ever since a most intriguing conference presentation by the authors several years ago discussing the problems of establishing 'quality metrics' for iris recognition. This chapter reports on the quality and performance data developed in the (U.S.) National Institute of Standards and Technology (NIST) 2006 'Iris Challenge Evaluation' (ICE) from algorithms submitted by three different groups. A primary impediment to any discussion of quality is the creation of an unambiguous, non-tautological definition. Chapter 5 uses a definition previously developed by NIST for fingerprint quality (NISTIR 7151, 2004), which has always struck me as clear and pragmatic, as illustrated in the paragraph 'The underlying premise for quality measures is that they are predictive of performance. In addition, a desirable property of the quality measures is that they are universal. A quality measure is universal if it is predictive of performance for a class of algorithms'. Based on the agencies submitting to ICE, it is reasonable to guess that all three algorithms use a binary encoding of local phase information and thus represent a single 'class' of algorithm. However, Chapter 5 reports that no correlation is observed in quality metrics across submissions and little correlation is observed between quality measure and performance within submissions by the same agency. This demonstrates that, at least in 2006, no 'universal' iris quality metric was available and much work remained to be done even on vendor-specific quality determination. Lacking such work, statements like 'We need to be collecting high quality images' are vacuous, which is the motivation for Chapter 4, 'Iris Quality Metrics for Adaptive Authentication'. Chapter 4, although going on to make excellent progress towards resolving the above issues, is less clear regarding a workable understanding of 'quality'. 'The main role of these (quality) metrics is to quantify, at the stage of data acquisition or a later processing stage, what information an iris image contains to discriminate the iris class, which it represents, from all other iris classes in a database'. 'What information an iris image contains to discriminate' cannot be quantified until we determine what information is needed to discriminate. Without reference to an algorithm, analysed to determine what information in the iris class will be used for successful comparisons, quality metrics lack meaning. This chapter uses a single 'Gabor filter-based' algorithm, with the implication that it is of the Daugman type, employing binary encoding of local phase information. Even if we cannot assume that results will be representative of all algorithms of this type, at least the chapter supplies evidence of metrics that do correlate well with system performance: blur, interlacing, segmentation, illumination, dilation, off-angle, pixel count and occlusion. For all but occlusion, their impact on the ROC is given. The ROC-based method of demonstrating these results parallels NIST work on fingerprint quality and is certainly much clearer to me than the 'pruning' method of Chapter 5. Chapter 19, 'Optics of Iris Imaging', is a beautifully written chapter that should serve as a model for anyone asked to contribute to an edited scientific work like this one. This is the first paper I have seen on the optics of the iris imaging process – the bulk of scientific papers over the years focusing instead on the development of comparison and segmentation algorithms and error rate estimation. This chapter gives a clear overview of basic optics, followed by a discussion of the relationship between lens characteristics, illumination wavelength, exposure time, sensor pixel size and signal-to-noise ratio. The impact of engineering design trade-offs on the various quality metrics, suggested in Chapter 4, implies that the answer to the question, 'What constitutes quality?', which from Chapter 5 appears to be algorithm dependent, will have a strong effect on both the cost and usability of an iris capture system. So it turns out that capture hardware, illumination wavelength, algorithm and error rates are more closely linked than I had previously imagined. Although the Handbook covers a broad range of technology issues, lacking is any attempt to explore the details of the Daugman algorithm. Complex Gabor filters require four parameters: amplitude, frequency, direction and decay. Phase encoding renders amplitude irrelevant, but how should the other parameters be chosen? With the 2011 expiration of the 1994 Daugman patent, perhaps there will be another flood of research with commercial motivations, this time on parameter selection. Also missing from the Handbook is discussion of the 'best of n' selection of the best Hamming distance after n shiftings of the iris code, as introduced by Prof. Daugman to accommodate small variations of in-plane angle of the iris in mated image pairs. This rotation degrades the non-mated score distribution in favour of the mated-score distribution, decreasing false negatives but increasing false positives, and providing an overlooked link between at least one factor in acquisition 'quality' and the non-mated score distribution. When present, this correction accounts for the asymmetry of the 'non-match' distribution as seen in Figure 11.4, which would otherwise be symmetric, as in the figures of Chapter 8, 'Visible Wavelength Iris Recognition'. The lack of correction in Chapter 8 allows the author to speculate on visible illumination false match rates of 1 in 1012 at HD = 0.33, a value which is not achieved even within 6 orders of magnitude by near infrared imaging systems using 'best of n', as previously discussed. Determining the impact of such a tunable correction under various data collection conditions across illumination wavelengths would be a valuable contribution. More importantly, perhaps, is the notable lack of discussion in any of the chapters on human factors and other interface design issues. Although human factors experts were among the first to scientifically study biometrics (for example, the late Prof. Gary Poock from the Naval Postgraduate School in Monterey, CA), our field has gotten away from such fundamental considerations. The use of databases of stored iris images, culled to remove 'poor quality' images as discussed in several chapters, avoids important questions of the effect of system design on data subject experience, public acceptance and error rates. Further, a case study on an operational system, including a Return-on-Investment analysis, would have been most welcome. I was pleased to see no speculative attempts to predict the future, but felt that insufficient space was given to seemingly prescient current laboratory systems, such as iris imaging of persons walking or at a distance. Although there were secondary mentions of 'Iris on the Move' technology in Chapter 8, 'Visible Wavelength Iris Recognition', Chapter 12, 'Fusion of Face and Iris Biometrics', Chapter 13, 'Methods for Iris Segmentation', and Chapter 14, 'Segmentation of Periocular Images', and discussion of data acquisition to distances of 8 m in Chapter 8 and 3 m in Chapter 19, 'Optics of Iris Imaging Systems', both topics – imaging with motion and distance – could have warranted their own chapters. More space devoted to the challenges of longer range iris recognition of unaware and moving data subjects might also serve to warn or console us as to its technical (in)feasibility and potential privacy implications. In fact, privacy issues are not discussed in the Handbook at all, save for a single mention in Chapter 2 which left me wondering about the difference between privacy and security. Accepting the claim in the Forward that iris recognition is far more accurate than face recognition, it could be argued that iris acquisition at a distance of unaware subjects could present privacy concerns specific to this technology. Chapter 16, 'Introduction to IrisCode Theory', although perhaps the least accessible, is special to me for a very personal reason. Prof. Richard Hamming was a pioneer in digital signal processing and one of the great and memorable characters at the Naval Postgraduate School in the latter part of the twentieth century. He kindly agreed to deliver the Keynote Address at the Eighth Biometric Consortium Conference held at San José State University in 1996, which may have been his last public lecture. Prof. Hamming always said that his goal was to have the word 'hamming', when used as an adjective, written without an initial capital. He pointed out that 'euclidean distance' is frequently written this way. Although the term appears many times in the Handbook, only Chapter 16 uses the word 'hamming' without the capital. Perhaps, the authors of Chapter 16 have heard me tell this story before, or perhaps they came up with the idea on their own. In either case, I know that Prof. Hamming is now having quite a chuckle, having gotten his final wish. We are left with the clear conclusion that iris recognition technology has made great strides in the laboratory when evaluated against databases of images. We can acquire images at many wavelengths, segment and store those images, measure their quality and compare them in multiple ways. If additional information is needed for human individuation, we can add face or periocular data. We can secure systems or attack them. The next phase for the technology will be a full analysis of operational deployments, including detailed examination of human factors issues. How do we design hardware and software systems to collect good quality images from people of all ages and abilities? What are the limits to iris acquisition at a distance of aware or unaware persons? How do we assess Return on Investment? What are the privacy implications peculiar to iris recognition? Overall, this is a book well balanced and well worth having in one's professional library, even at an on-line, hardcover price of £79.95. With this clear labour of love, Burge and Bowyer have made an outstanding contribution to our science of automated human recognition, for which they and the contributors are to be strongly congratulated. Review by JamesWayman, San Jose State University, USA