CCRNet: a novel data-driven approach to improve cross-domain Iris recognition

Abstract

In spite of the prominence and robustness of iris recognition systems, iris images acquisition using heterogeneous cameras/sensors, is the prime concern in deploying them for wide-scale applications. The textural qualities of iris samples (images) captured through distinct sensors substantially differ due to the differences in illumination and the underlying hardware that yields intra-class variation within the iris dataset. This paper examines three miscellaneous configurations of convolution and residual blocks to improve cross-domain iris recognition. Further, the finest architecture amongst three is identified by the Friedman test, where the statistical differences in proposed architectures are identified based on the outcomes of Nemeny and Bonferroni-Dunn tests. The quantitative performances of these architectures are perceived on several experiments simulated on two iris datasets; ND-CrossSensor-Iris-2013 and ND-iris-0405. The finest model is referred to as "Collaborative Convolutional Residual Network (CCRNet)" and is further examined on several experiments prepared in similar and cross-domains. Results depict that least two error rates reported by CCRNet are 1.06% and 1.21% that enhances the benchmark for the state of the arts. This is due to fast convergence and rapid weights updation achieved from convolution and residual connections, respectively. It helps in recognizing the micro-patterns existing within the iris region and results in better feature discrimination among large numbers of iris subjects.

Deep Learning for Iris Recognition: A Review

“…Some studies introduce residual connections when designing network structures for IR tasks. [53] uses the residuals connection from ResNet. Since residual connections allow the network to be built deeper without degrading performance.…”

“…Since residual connections allow the network to be built deeper without degrading performance. The model used in [53] is the Collaborative Convolutional Residual Network (CCRNet). Experimental results show that CCRNet has good performance with an EER of 1.06% and 1.21% for ND-CrossSensor-Iris-2013 and ND-IRIS-0405.…”