Advanced Comparison Techniques for Challenging Iris Images

Abstract

While better segmentation is certainly a highly effective approach to target iris processing for less constrained images, Daugman’s rubbersheet transform model presents a rather simplified anatomical model of pupillary dilation. Therefore, it is likely that two normalized iris images are not perfectly aligned. The majority of feature extraction approaches extracts binary output (iris-codes) from the obtained normalized textures [>44], and employs the fractional HD over different bit shifts between iris-codes in order to determine a degree of similarity at comparison stage. By shifting one of the two iris textures to be compared, or its corresponding iris-code, and calculating a comparison score employing the HD for each shift position, it is possible to account for the optimal alignment of both textures or iris-codes and to achieve so-called rotational invariance, i.e. likely present rotations of the head and consequently iris images are tolerated. However, the polar unwrapping with its simplifying assumptions may cause irrevocable mapping distortions even in case of generalizing from circular or elliptic shapes to AC-based approaches. In this case it is desirable to employ more sophisticated comparison techniques in order to exploit all the available information. Indeed, very few studies have proposed new or compared different binary similarity and distance measures, and it is a common agreement that HD-based comparison as proposed by Daugman [116] is the best method for this task. However, especially for unconstrained imagery more sophisticated comparison techniques are an efficient means to increase recognition accuracy without the necessity of re-enrollment. Another aspect to consider is speed-up of the recognition process. Traditional identification mode assessment involves a 1:n comparison, where n is the number of registered gallery subjects, and consequently does not scale well with respect to the number of enrolled users. Especially the ongoing Aadhaar project in India employing fingerprint and iris biometrics to uniquely identify each Indian citizen, and the necessary de-duplication checks during enrollment to avoid the issue of multiple Aadhaar numbers to a single individual make efficient identification an ultimate goal. Indeed, by employing partial matching and indexing techniques, huge amounts of processing time can be saved [270]. While traditional approaches typically sacrifice recognition accuracy in favor of more efficient comparison, there are methods to increase both accuracy and speed at the same time by effectively re-arranging iris-codes according to bit-reliability [440]. A similar technique may also be used to exploit single-instance multi-algorithm iris biometrics to increase accuracy without extending the total template size [439]. Therefore, the second part of this chapter targets the problem of iris recognition under unideal conditions from a different perspective, namely the application of sophisticated iris comparators and advanced serial iris identification techniques. Table 9.1 lists some of the techniques in a comparative manner accounting for computational cost, accuracy, and needed enrollment samples. Especially the latter is a typical restrictive condition in many application scenarios (e.g., forensics, high-throughput enrollment), where only a single enrollment sample is available.