Iris Recognition for Biometrics Based on CNN with Super-resolution GAN

Abstract

Biometric technologies can realize high-security systems. Although camera performance and photographic environment cause a low signal-to-noise ratio, super-resolution (SR) techniques, such as generative adversarial networks (SRGANs) and deep convolutional neural networks (DCNNs), can improve image quality. This study aimed to investigate the effect of the SRGANs on individual identification by DCNNs, assuming external image noise and a prefiltering process in actual cases of iris recognition. After downgraded iris images were improved by the SRGANs, a DCNN classifier predicted the individuals from the restored images. The accuracies of the DCNN classifier were higher in the SR images using the Bicubic method or squared mean errors than the SRGANs focusing on perceptual loss. This result suggests that it may be easier for the DCNN classifier to create image features based on the pixel-based differences (i.e., high peak signal-to-noise ratio), rather than on the perceptual image differences. In the future, a robust security system based on SR methods may be capable of assessing a health condition using the images obtained for individual certification.