Iris patch texture detection algorithm based on residual fusion attention and U-shaped network
Bo Zhang, Xiaoyun Mei, Changpeng Wang · 2023
In the detection of iris plaque, traditional machine learning algorithm is unable to detect any location, size and iris plaques interfered by light spots. In this paper, an DIPS-Unet model for iris patch texture detection is proposed. In the encoder stage, a residual convolution fusion attention module is introduced to obtain deep semantic information and focus on details. In the decoder stage, quadratic linear interpolation and deconvolution are used successively. Finally, cross entropy and Dice mixing loss were introduced to solve the class unbalance problem. At the same time, the preprocessing algorithm is proposed and the mirror symmetry method is used to extend the data set. Experimental results show that the proposed method solves the influence factors of traditional algorithms, and the deep learning model is applied to detect iris pigment texture for the first time. Compared with the original model, the intersection ratio of the improved model is increased by 9.98%, F1 score is increased by 0.0632, and accuracy rate is increased by 7.14%.