基于特征关联性的人脸高层特征研究

Research on high level features of face based on feature correlation

在特殊应用领域，注册者只能注册一张人脸信息，使得人脸注册信息极为有限，给人脸识别带来很大的限制。文章以参考集为基础，开展了基于特征关联性的人脸高层特征研究，通过对参考集中该人脸的所有图片特征均值和训练集、测试集中数据进行距离计算，将对应训练集、测试集中的各人脸的距离依次组合构成向量作为该脸的高层特征，该方法在很大程度上解决了注册信息缺失的问题；在Multi-PIE库和扩展YaleB库中进行了实验，并与基于稀疏表示的分类（sparse representation-based classifier，SRC）算法进行了对比。实验表明：该算法比余弦距离分类方法人脸识别的正确率提高5％～6％；与SRC算法相比，该算法更具有优越性。研究结果对单训练样本条件下的人脸识别研究有一定作用。

In special cases the registered person can only provide one picture, thus face registration information is extremely limited, bringing a lot of restrictions to face recognition. In this paper, a high level feature of human face based on the correlation of features is studied by using reference data. For a certain face in the training data or test data, the feature distances between it and all the faces in the reference data are calculated, and these distance values are combined as a high level feature. This method solves the problem of incomplete registration information to a great extent. The results of the experiments on Multi-PIE database and the extended Yale face database B indicate that the recognition accuracy of the proposed algorithm rises by 5 %-6 %comparing to that of the cosine distance classification method. Comparing with the sparse representation-based classifier（SRC） algorithm, the performance of the proposed algorithm has superiority. This study is valuable for promoting the research on face recognition under the condition of single training sample.