Abstract:

Single sample face recognition has become a hot topic in the field of face recognition since its wide application in real life. The lack of training samples and the dramatic inter-class variations of facial expression, illumination, and occlusion make it difficult to study. The traditional face recognition method based on sparse representation needs a large number of training samples to construct an over-complete dictionary, so the recognition accuracy is significantly dropped under the single sample condition. To solve this problem, a supervised autoencoder based method is proposed to generate faces with variations, which can automatically generate face images with variations while preserving identity information for dictionary expansion under the condition of a single sample. To a certain extent, this method can alleviate the problem of under sampling under the condition of a single sample, and make up for the difference of face variance information between the training set and test set, making the traditional sparse representation method suitable for single sample face recognition. Experimental results on public databases not only prove the effectiveness of the method but also show strong robustness to different face variations in the test set.

Key words: single sample face recognition, supervised auto-encoder, sparse representation, dictionary learning

摘要：

单样本人脸识别因其在现实生活中的广泛应用而成为人脸识别领域的热门话题。单张训练样本条件下训练样本的缺少和复杂的类内人脸表情、光照、遮挡变化给单样本人脸识别研究带来困难。传统的基于稀疏表示的人脸识别方法需要大量的训练样本构成过完备的字典，因而在单样本条件下识别效果明显下滑。针对这一问题，提出一种基于有监督自编码器的带变化人脸样本生成方法，在保留身份信息的同时自动生成带变化的人脸图像用于单样本条件下的字典扩充，一定程度上缓解了单样本条件下的欠采样问题，弥补了训练集和测试集间的人脸变化信息差异，使得传统的稀疏表示方法能够适用于单样本人脸识别问题。在公共数据库上的实验结果不仅证明了该方法的有效性，而且对测试集中不同的人脸变化也展现出了较强的鲁棒性。

关键词: 单样本人脸识别, 有监督自编码器, 稀疏表示, 字典学习