Abstract: Most face images taken from natural environment have occlusion, which has always been a huge challenge for face recognition. The mainstream deep model used for face recognition does not have particularly good identification performance for occlusion of face images. In order to solve the problem that the recognition rate decreases because occlusion exists and the occlusion position is uncertain for deep model, this paper proposes a face recognition algorithm of occlusion location based on PCANet (principal components analysis network), which combines deep learning and feature point occlusion detection. Classifier is used for key point detection, and PCANet deep learning model is used for feature extraction to form SVM (support vector machine) training model group. The occlusion discrimination classifier locates occlusion, combines the group of feature model to complete the occlusion face recognition task, and has strong robustness to facial expression changes. The experimental results show that the algorithm has achieved very good results for common occlusion types, and the extreme types of large-area occlusion also have a high recognition rate.

Key words: deep learning, key points, occlusion discrimination, group of model, face recognition

摘要： 自然环境中的人脸图像大部分带有遮挡，这对于人脸识别一直是巨大的挑战，用于人脸识别的主流深度模型对于遮挡人脸图片并不具有特别好的识别性能。针对深度模型由于遮挡的存在以及遮挡位置不确定所导致的识别率下降的问题，提出一种结合深度学习和特征点遮挡检测的PCANet下的遮挡定位人脸识别算法。分类器用于关键点检测，使用PCANet深度学习模型进行特征提取，形成支持向量机（SVM）训练模型组。遮挡判别分类器定位遮挡，结合特征模型组完成有遮挡人脸识别任务，并且对于表情变化有很强的鲁棒性。实验结果表明，该算法对于常见遮挡类型取得了非常好的效果，对于大面积遮挡的极端类型也具有很高的识别率。

关键词: 深度学习, 关键点, 遮挡判别, 模型组, 人脸识别