DADCNet: Dual attention densely connected network for more accurate real iris region segmentation

Most existing performance evaluation standards for iris segmentation algorithms, such as the typical recall, precision, and F-measure (RPF-measure) protocol, are based on a pixel-to-pixel comparison between the mask image obtained after segmentation and the corresponding ground truth (GT) image. However, one of the most important problems is that if the published GT images have errors when locating the iris region, then the reference value of these evaluation indicators will be reduced, which is not conducive to the development of iris recognition technology. To address this problem, this paper proposes to use a mask image segmented by a deep learning method to replace the corresponding GT image. The main work of this paper is as follows. First, a dual attention densely connected network (DADCNet) containing two attention modules and an improved skip connection is proposed to segment the real iris region more accurately than the corresponding GT image. Second, the recognition performance of the two input classes obtained from the mask image after DADCNet segmentation and the corresponding published GT image in the same recognition network is utilized to further show that the former is more reliable in positioning the real iris than the latter. To make the proposed network more convincing, extensive experiments are conducted on four representative and challenging iris databases, which is obtained under different spectral conditions. These results show that the proposed DADCNet achieves state-of-the-art performance and that the mask image obtained after DADCNet segmentation can replace the published corresponding GT image.