Study Of Face Recognition Based On Deep Learning

Face recognition is one of the objects in pattern recognition field, and it has been widely applied in security, finance, judicature, network transmission and so on. Although human face has rich invariant features, there are several factors affecting face recognition because of external disturbance. Hence, it is a key and challenge to get intrinsic features which can be invariant to external noises. At present, most facial features are human-crafted and are strongly dependent to specific problems, and are not easily extended. Therefore, it is a pervasive problem to be solved that design a kind of feature which is universal and less human intervening.Aiming to the problem of feature extraction, we propose an unsupervised feature learning method called Multi-scale Convolutional auto-encoder based on the idea of deep learning. MSCAE sets multi-scale convolution kernels in convolution layers to extract multi-scale features, which reflect facial natural contents, and can better restore human face. Its framework is a hierarchy of alternating filtering and subsampling, which makes features be invariant to rotation, translation, scale and other form of deformation. Training of MSCAE utilizes unlabeled samples to proceed encoding and reconstruction, and errors between input and reconstruction are minimized to adjust parameters. Intermediate codes of former layer are used as input of next layer to train this layer with the same way. After all layers being trained, freeze the reconstruction layers and stack all the encoding layers to extract features for classification. Experimental results with Neural network(NN) on ORL and Yale face datasets suggest that multi-scale features are superior to single-scale ones on recognition rate and efficiency. Furthermore, fusion features of MSCAE and HOG(Histograms of Oriented Gradients) get higher recognition rate than either of them.Besides, aiming at layer-wise training mechanism of deep learning, convolutional auto-encoders are applied to pre-train each layer of convolutional neural network without supervision, providing stable filters. Then labeled samples are used to fine-tune the whole networks supervised and classify human faces at the same time. Experimental results on ORL database suggest that this kind of unsupervised pre-training can accelerate the rate of convergence. CNN with little labeled samples can get high recognition rate.Features extracted by MSCAE get high recognition rate in face recognition. But MSCAE uses raw pixel as input to extract features, which makes it sensible to illumination. Therefore, future work can be carried out to improve the algorithm specific to this problem. Besides, character of MSCAE that restricts output equal to input can be used to deal with pose and other problems.