Influence of Segmentation on Iris Images in the Visible Spectrum Using Deep Multi-Task Attention Network

Abstract

The iris recognition is not benefited significantly despite the growth of Deep Learning (DL) in several fields of image processing and computer vision.  Mostly the emerging researches on iris recognition have been focused on new approaches to generate robust and discriminative iris representations based on methods akin to classical iris recognition pipelines. The images of iris captured in the non-cooperative areas often suffer from adverse noise that can challenge most of the earlier iris recognition techniques. To overcome this issue, Influence of Segmentation on Iris images in the visible spectrum using Deep Multi-Task attention network is presented in this research. The iris mask, boundaries of outer iris and parameterized inner are jointly attained while modeling them actively into a unified multi-task network.  The segmentation is initiated by pupil through the state of the art circular Hough transform method. Then the iris-sclera boundary will be estimated by a very thin slab in the actual image. In this work SBVPI and CASIA Thousand datasets are utilized for experiments and many interesting findings will be reported. The combined approach will perform best compared to all trained models.  How the segmentation accuracy influence the performance of iris recognition is evaluated and also examined that if segmentation will be needed at all.