Research On Face Recognition Algorithm Based On Convolutional Neural Network

Due to the continuous progress of deep learning,it has greatly improved the performance of various computer vision technologies.Face recognition,as one of the most extensive computer vision technologies,develops rapidly and has broad application prospects in the military,financial,e-commerce,security and defense,and other fields.The early research on face recognition is based on the features of the manual design,which is difficult and requires strong prior knowledge.It is difficult to design effective facial features.As one of the deep learning algorithms,a convolutional neural network has the natural ability to automatically extract features and has become the mainstream direction of face recognition.However,because of non-rigid of face,it is different from other image recognition and is greatly affected by changes in illumination,shielding,and posture.How to set up reasonable convolutional neural network structure,loss function and other problems to extract more effective features for face recognition has become the focus of dissertation.The contents of this dissertation are as follows:(1)The research process of deep learning is reviewed and the development status of face recognition is introduced.In this dissertation,the layers of convolutional neural network theory are introduced in detail,and the back-propagation algorithm,residual network and Softmax loss function are analyzed in detail.(2)Aiming at the problem that angular Softmax loss exists in strong constraint,an algorithm of angular triple Softmax loss is proposed to fine turn the angular Softmax loss pre-training model.Firstly,the proposed algorithm improves the original convolutional neural network structure by adding1×1convolution kernel and pooling layer between different residual blocks to select more effective features.Then,the angular triple loss is used to fine-tune the pre-training model to reduce the strong constraint conditions of the difficult samples.At last,in the test,the original face image features and the horizontal face image features are extracted,and the two features are added as the final facial feature expression,so as to enrich face feature information and improve face recognition performance.The experimental results show that in large-scale face recognition,the proposed algorithm can improve the face recognition rate only with a single model and a relatively small training set.(3)Aiming at the shortcoming of feature separability of Softmax loss,a discriminative face recognition algorithm based on deep convolutional neural network is proposed.Firstly,according to Softmax loss feature distribution,an intra-class cosine similarity loss between feature and weight vector to make the intra-class more compact and the inter-class as separate as possible.Then,on the basis of Softmax loss,normalized features are used to better simulate low-quality face images,and normalized weights were used to reduce the class imbalance,making it consistent with cosine similarity measurement during testing.Finally,combined the normalized Softmax loss and the similarity loss of intra-class cosine are fine-tuned on the pre-training model.The experimental results show that the algorithm improves the discrimination of features,enhances the generalization ability of the model,and can improve the face recognition rate.