COMPUTATIONALLY EFFICIENT SERIAL COMBINATION OF ROTATION-INVARIANT AND ROTATION COMPENSATING IRIS RECOGNITION ALGORITHMS
Abstract:
Rotation compensation is one of the computational bottlenecks in large scale iris-based identification schemes, since a significant amount of Hamming distance computations is required in a single match due to the necessary shifting of the iris codes to compensate for eye tilt. To cope with this problem, a serial classifier combination approach is proposed for iris-based identification, combining rotation-invariant pre-selection with a traditional rotation compensating iris code-based scheme. The primary aim, a reduction of computational complexity, can easily be met - at comparable recognition accuracy, the computational effort required is reduced to 20% or even less of the fully fledged iris code based scheme. As a by-product, the recognition accuracy is shown to be additionally improved in open-set scenarios.
1 INTRODUCTION
Iris recognition technology has been dominated over years by the commercially successful algorithm of J. Daugman (Daugman, 2004). This algorithm basically extracts local iris features from polar iris images by convolution with 2-dimensional complex Gabor atoms, quantizing the resulting phase information into 2 bits per coefficient. The basic idea of extracting local intensity variations from iris texture has been followed employing other types of transforms and methods as well, e.g. in the spatial domain or in the wavelet domain. All these approaches share the property of being sensitive against eye tilt, i.e. they are intrinsically not rotation invariant due to the usage of local spatial information. Therefore, in order to compensate potential rotation, in all these algorithms the templates in the matching process are shifted against each other for a certain amount, and taking the minimal template distance among all shifted versions as the actual distance. Obviously, depending on the amount of shift that is required for a certain application (i.e. the amount of rotation that is to be expected), these operations may amount to a significant number of matching operations performed, which can become prohibitive in an identification scenario. Rotation-invariant iris features therefore represent an attractive alternative. Due to the significant computational demand associated with transform domain processing, spatial domain techniques working directly on the iris texture are of specific interest in our context. Du et al. (Du et al., 2006) employ first order moments of the iris texture line-histograms. While this technique is successful in providing rotation invariance and consequently fast matching procedures independent of the eye’s position, it fails in terms of recognition accuracy. This is where our approach comes in. In this work we combine a spatially-based rotation invariant iris recognition approach with a traditional local-feature based scheme into a serial classifier combination. The aim is to result in reduced overall computational demand as compared to classical rotation compensating schemes while at least maintaining their recognition accuracy. This is achieved by using the first scheme to determine a certain amount of the highest matching ranks of the entire database (this can be done quickly due to the high speed of the first

Download full report
http://cosy.sbg.ac.at/~uhl/VISAPP.pdf