基于深度自编码网络的异质人脸识别

HETEROGENEOUS FACE RECOGNITION BASED ON DEEP AUTO-ENCODER NETWORKS

针对异质人脸识别中对不同模态数据间关系建模的问题,提出一种基于深度自编码网络的异质人脸特征提取和识别方法。首先用一个深度降噪自编码网络从两类异质人脸图像中提取人脸的高阶特征,并通过类别监督信号产生的目标函数来对网络进行微调,最后利用最近邻分类器对已提取特征分类,完成异质图像间的匹配。在CUHK、AR、CASIA HFB、SVHN与MNIST数据集上的实验结果表明,与目前基于子空间学习的异质人脸识别方法相比,该方法取得了更高的识别率,并且在基于异质图像的数字识别上表现出一定优势。

Considering the problem of modelling the relation between different modalities data in heterogeneous face recognition,we proposed a heterogeneous face feature extraction and recognition method,which is based on deep auto-encoder networks. The method first extracts the high-order feature from two kinds of heterogeneous face image respectively using one deep denoising auto-encoder networks.Then,it fine-tunes the network through the objective function generated by category monitoring signals. Finally,it uses nearest neighbour classifier to classify the extracted features so as to complete the matching of heterogeneous face images. Results of experiments conducted on datasets of CUHK,AR,CASIA HFB,SVHN and MNIST showed that,compared with existing subspace learning-based heterogeneous face recognition methods,the proposed one reaches higher recognition rate,and exhibits certain advantage in heterogeneous image-based digital recognition.