基于轻量化网络的眼部特征分割方法

针对高分辨率眼部图像的瞳孔、虹膜特征快速识别与检测问题，提出了一种轻量化语义分割网络DIA-UNet （double input attention UNet）.它采用对称双编码结构同步获取眼部灰度图及其轮廓图特征，并通过双注意力机制实现了解码端的特征筛选，将深层融合特征作为语义分割输出.在CASIA-Iris-Interval和高分辨率瞳孔数据集上测试结果表明，与其他轻量化语义分割网络相比，本文提出的DIA-UNet在保证虹膜、瞳孔分割准确率的同时网络参数个数仅有0.076 Million，处理速度高达123.5 FPS. 

A Lightweight-Network-Based Segmentation Method for Eye Features

To deal with the problem of accurate and fast recognition and detection of pupil and iris features with high-resolution eye images, a lightweight semantic segmentation network DIA-UNet (dual input attention UNet) was proposed based on the UNet framework. It was arranged that, a symmetric dual coding structure was adopted to obtain the features of the eye grayscale image and its contour image synchronously, the dual attention mechanism was used to carry out the feature filtering on the decoding end, and taking the deep fusion features as a semantic segmentation output. The test results from the CASIA-Iris-Interval and high-resolution pupil datasets show that, compared with other lightweight semantic segmentation networks, the proposed DIA-UNet can guarantee the accuracy of iris and pupil segmentation, while the number of network parameters is only 0.076 Million and the processing speed is up to 123.5 FPS.