Research And Implementation Of Face Recognition Based On Convolution Neural Network

Nowadays, with the rapid development of the related theories and applications of computer vision, the superiority of the application of computer vision technology in daily life is becoming more and more important. This thesis mainly studies the application of the convolution neural network model,which is belong to deep learning method,in the field of human face recognition in natural scenes. Compared with the traditional face recognition method, the deep Convolutional Neural Network model(CNN) does not need to design the feature extraction algorithm, which is complex and time-consuming, it just have to design an effective neural network model, and the model learn from a large number of training samples by an end to end training, then this method can reach an good classification accuracy. The performance and effectiveness of the method are mainly determined by the design of the model structure, so the key point of this thesis is to design a reasonable neural network model, and some related technologies is also applied in the model to ensure that the model can converge on the training set quickly and stably.The main contents of this thesis include:(1) In this thesis, some basic theories of the convolutional neural network are summarized. Convolutional neural network developed from the traditional neural network, so the network structure of traditional neural network, gradient descent method and BP algorithm(Error Back Propagation Algorithm) are described. And then transition to the description of the related theories of convolutional neural network, such as convolutional layer, pooling layer etc. Finally, this thesis illustrates the general structure of convolution neural network model by introducing the classic LeNet-5 network.(2) By reducing the number of parameters in the raw VGG convolutional neural network reasonably, an improved Lightened VGG network model has been designed, and a new parameter initialization method is applied in this model, which is better than randomly parameter initialization method, to reduce the time of model convergence. At last, this new model not only solves some issues which had occurred in the original model, such as higher-quality hardware requirements, the difficult of training, and so on, but also is successfully applied to face recognition in natural scene, which reached 94% accuracy rate on the strictly pre-processed LFW(Labeled Faces in the Wild) dataset. Then, to further improve the ability of the model to extract the features of more complex images, a Siamese model is used and illustrated in detail.(3) In this thesis, a residual convolutional neural network also is designed by applying a new residual learning theory. The layers of this model reached to 34. To solve the difficult of convergence in this model, a new parameter initialization method is used, and Batch Normalization technique is applied to make the model more stable. According to the result on LFW, the accuracy rate of this model can reached 96%, which is better than the Lightened VGG model.(4) Finally, a face recognition system based on the Real-Time surveillance video in real scenario is implemented by applying the models mentioned before. The function and process of each module in the system are illustrated in detail, and the accuracy of the test, which is carried out on a self-built face database, is 93%. The system verifies the effectiveness of this method, and it can meet the requirements of face recognition applications in the surveillance video.