基于循环生成对抗网络的含遮挡人脸识别

人脸图像的遮挡会严重影响人脸识别准确率，当前处理带遮挡人脸识别的方法主要有丢弃法和修复法两种。丢弃法因忽略或丢弃大量遮挡区域的有效特征易造成识别准确率不高，而当前大多数修复法需要原图的相关信息，限制了其应用。针对现有含遮挡人脸识别方法存在的问题，提出一种基于循环生成对抗网络的人脸识别算法，通过利用2对生成器和判别器进行循环训练，实现遮挡人脸图像的盲修复，此过程不需要除遮挡区域外的原图信息。在此基础上，采用ResNet-50网络对修复后人脸进行识别，该网络通过跳跃连接消除深层卷积神经网络中的网络退化问题，能够降低网络训练的难度，且不会增加额外的参数和计算量。针对盲修复后人脸特征存在类内差异大和类间差异小的特性，引入一种能够量化类间距离的分类损失函数RegularFace作为识别网络损失函数。实验结果表明，与DCGAN+CNN算法相比，所提算法对不同遮挡类型和遮挡区域的人脸图像识别准确率均有所提高，当线性遮挡面积为40%时，所提算法的识别准确率提高了14.4个百分点。

Face Recognition with Occlusion Based on Cyclic Generative Adversarial Networks

Face image occlusion has an important impact on the accuracy of face recognition.Currently, two primary methods are used to process face recognition with occlusion:discard and repair.The discarding method ignores (discards) the effective features of several occluded areas, which easily leads to a low recognition accuracy.However, most of the current restoration methods require a known original image with related information for restoration, which limits its application.Aiming at the problems existing in the existing face recognition methods with occlusion, this paper proposes a face recognition method based on cyclic Generative Adversarial Networks(GAN).This method first uses two pairs of generators and discriminators to cyclically train to realize the blind restoration of the occluded face image.This process does not require information of the original image, except the occluded area.On this basis, a ResNet-50 network recognizes the repaired face after, eliminating the network degradation problem in deep convolutional neural networks using jump connections, which can reduce the difficulty of network training without additional parameters and computation.Considering the large and small inter-class differences in face features after blind reparation, a classification loss function that can quantify the distance between classes is introduced as the recognition network loss function.Compared with the DCGAN+CNN algorithm, the results show that, the recognition accuracy of the proposed algorithm improved for face images with different occlusion types and occlusion areas, and, when the linear occlusion area is 40%, the recognition accuracy of the proposed algorithm improved by 14.4 percentage points.