Two-Stage and Two-Discriminator generative adversarial network for the inpainting of irregularly incomplete iris images

Abstract

Due to the influence of the light source environment during image acquisition or the subject not fully opening their eyes, there are phenomena such as light spot interference, eyelash or eyelid occlusion in the iris area. This will cause the loss of effective iris information, ultimately affecting the success rate of recognition. To address the aforementioned issues, this paper introduces a Two-stage and Two-discriminator Generative Adversarial Network (TT-GAN) to restore irregularly incomplete iris images. The specific process is as follows: Firstly, TT-GAN employs gated convolution in place of ordinary convolution to more effectively extract image information; Secondly, TT-GAN incorporates a PC-Attention attention mechanism within the generator network to enhance the network's restoration capabilities; lastly, to ensure consistency in global and local semantic information of the repaired images, dual SN-PatchGAN discriminators are utilized. The experimental results based on three different spectral iris datasets, IITD, CASIA-Iris-Interval, and UBIRIS.V2, show that the repaired iris images significantly improve objective evaluation indicators compared to the incomplete iris images. Among them, TAR increased by 93.42 %, 68.41 %, and 59.72 %, respectively; EER increased by 27.17 %, 30.18 %, and 23.78 %, respectively; PSNR increased by 17.8712 dB, 18.818 dB, and 24.6053 dB respectively; SSIM increased by 0.0325, 0.0529, and 0.1337 respectively.