Attention Skip Connection Dense Network for Accurate Iris Segmentation

Abstract

AbstractAs a key step in the iris recognition process, iris segmentation directly affects the accuracy of iris recognition. How to achieve accurate iris segmentation under various environmental conditions is a big challenge. This paper proposes attention skip connection dense network (ASCDNet), which adopts an codec structure, and uses dense blocks as a component of encoder to obtain richer iris features and alleviate the problem of gradient disappearance. In the improved skip connection, channel attention and spatial attention mechanisms are introduced to achieve effective information fusion through the connection of high and low layers. The experimental results on two iris datasets IITD, CASIA-Interval-V4 collected under near-infrared light and one iris dataset UBIRIS.V2 collected under visible light show that the proposed improved skip connection can effectively improve the performance of iris segmentation, the accuracy rates of segmentation are as high as 0.9882, 0.9904, 0.9941, respectively, outperforming most state-of-the-art iris segmentation networks.KeywordsIris segmentationASCDNetAttention mechanisms