Face Recognition Based On Deep Learning

The traditional face recognition pipeline uses shallow structure to extract the feature of face image, which has limited ability to extract the connotation feature of face image, and the recognition result is not satisfactory. With the development of cognitive science and brain science, it provides a new perspective and pipeline for deep network technology. People simulate the structure of the hierarchical representation of the human brain, and put forward a variety of deep network models. The deep network adopts the structure of nonlinear and multi-layer network, and extracts the feature of face image layer-by-layer, which makes extracted features more abstract and discriminative, and is more suitable for complex classification problems. It has attracted the attention of many researchers and has become a new star in the field of pattern recognition, since the deep learning has been proposed. In this thesis, the deep learning technology be applied to face recognition, the main works are as follows.1. Face pose estimation has been widely used in face recognition and human-computer interaction. Because the gradient feature can well describe the difference of the attitude images, the combination feature between gray feature and gradient feature was taken as input to train the DBN network to achieve the classification of attitude. In this thesis, deep belief network with the three layers was constructed, and the parameters of network ware trained. Experiments ware done on the CAS-PEAL-R1 face database. The experimental results show that compared with the conventional method of only taking gray features as input for deep learning, taking the combination of gray features and gray difference features as input can achieve better performance.2. In this thesis, the advantages and disadvantages of image gray feature, LBP feature and gradient feature are analyzed, and three complementary fetures ware fused to extract more robust features. Firstly, the gray feature, LBP feature and gradient feature are extracted. And then three CNN networks ware constructed for reducing the dimensionality of gray feature, LBP feature and gradient feature, and extracting the more discriminative features respectively. Finally, the three discriminative features ware concatenated, and the dimensionality reduction and fusion ware carried out by the PCA algorithm. Cross-validation experiments ware done on the LFW face database. The experimental results show that compared with the feature extraction method based on single-feature and single-network, the feature construction method based on multi-feature and multi-network achieve better performance.3. In this thesis, a fine-turning algorithm of convolutional neural network is proposed, which reducing the variation of the intra-class features was taken as the optimizing target. The test experiments implemented on LFW face database and YTF face database which are not overlapping with the training set, and the equal error rate is increased by 6.50% and 4.04% respectively. The test results show that new algorithm can achieve better generalization performance.