基于深度自编码器的单样本人脸识别*

由于每个目标仅有一幅已知样本,无法描述目标的类内变化,诸多人脸识别算法在解决单样本人脸识别问题时识别性能较低.因此文中提出基于深度自编码器的单样本人脸识别算法.算法首先采用所有已知样本训练深度自编码器,得到广义深度自编码器,然后使用每个单样本目标的单个样本微调广义深度自编码器,得到特定类别的深度自编码器.识别时,将识别图像输入每个特定类别的深度自编码器,得到包含与测试图像相同类内变化的该类别的重构图像,使用重构图像训练Softmax回归模型,分类测试图像.在公共测试库上进行测试,并与其它算法在相同环境下进行对比,结果表明文中算法在获得更优识别率的同时,识别一幅图像所需平均时间更少.

One Sample per Person Face Recognition Based on Deep Autoencoder

Since there is only one sample for each subject, it is hard to describe intra-class variations of the subject. The performance of state-of-the-art face recognition algorithms declines in one sample per person(OSPP) face recognition. In this paper, an OSPP face recognition algorithm based on deep autoencoder(OSPP-DA) is proposed. In OSPP-DA, deep autoencoder is trained by all the images in the gallery firstly, and a generalized deep autoencoder(GDA) is generated. Then, the GDA is fine-tuned by the single sample of the subject, and a class-specified deep autoencoder(CDA) is obtained. For classification, query images are input to CDAs and the reconstruction samples of the corresponding subjects have the same intra-class variation as query images. A Softmax regression model is trained by the reconstruction samples and the query images are identified by the Softmax regression model. Experiments on public testing database are conducted and the results show the validity of OSPP-DA. Compared with some state-of-the-art algorithms, the proposed algorithm produces better performance with less time.