Dr-Srwgan: a Self-Trained Disentangle Representation and Super-Resolution Wasserstein Generative Adversarial Networks for Iris Segmentation and Occlusion Estimation

For research topic like image recognition using CNN, how to collect a large number of images for network training and testing is a common difficulty in practice. The same is true for model training for deep iris recognition. How to collect enough special case images to retrain the neural network is an important issue. The iris image itself is not easy to collect and requires special optical equipment, and the iris image under special conditions is more difficult to collect. In practice, special iris images often fail the iris recognition system (because of the failure in iris segmentation stage). Therefore, collecting a large number of special iris images will be of great help in training new deep iris recognition algorithms.In order to produce images that are sufficiently realistic and can be targeted for specific purposes, we propose a generative adversarial network with self-learning capabilities, called DR-SRWGAN. In this new GAN architecture, we comprehensively apply several recent GAN techniques, including: Pix2Pix, WGAN-GP, Super-Resolution GAN, and DFCN. After properly trained, DR-SRWGAN can randomly generate various special types of iris images and their precise Groundtruth labels according to the needs of the experimenter, including the inner and outer boundaries of the iris, and the iris mask, off-axis gazing angle, and a binary indicator for eye-glasses. In such way, it solves the problem of scarce data problem for deep iris recognition. The research results can also be applied to other areas including: iris recognition, object localization or semantic segmentation, high-precision eye tracking system, image analysis of ophthalmology-related diseases, iridology … etc.