The Research On Face Detection And Feature Extraction Algorithms In Face Recognition

As a model of biometrics,face recognition has a good application prospect in many areas of people's lives.Face recognition has made great progress in decades of research.At present,face recognition can achieve a relatively high accuracy under simple and fixed conditions.But face recognition under complex conditions is still a very challenging topics.Face detection and face feature extraction are easily affected by the background,light,attitude,occlusion,expression and other factors,resulting in a sharp decline in recognition rate.So it is very important to develop robust and efficient face detection and feature extraction algorithms.The main work is as follows:(1)This thesis introduces the background and significance of the research,research and analysis of the status quo of face recognition at home and abroad,elaborating the three main steps in the face recognition process.This thesis also introduces the relevant technologies such as machine learning model and depth learning model used in each step,and analyzes the existing problems in the mainstream algorithms.(2)Convolutional neural network(CNN)is effective for face detection in complex background.But in the process of detection,the number of windows brought by sliding window method is large,which makes the computation complexity of the algorithm relatively high.In the process of convolution neural network training,due to random initialization and the defect of activation function.So the convolution neural network convergence process caused by slow or non-convergence.A face detection algorithm based on Binarized Normed Gradients(BING)and convolutional neural networks is proposed,and a new activation function of CNN is proposed,and the corresponding initialization method is derived.Simulation results show that face detection algorithm based on BING and convolutional neural network has significantly improved training speed and detection speed,and has ensured high accuracy of face detection and robustness against complex background.(3)Face feature extraction method based on Garbor wavelet amplitude and phase has better robustness for illumination change.But the cascade feature fusion method of the face feature extraction method based on Garbor wavelet amplitude and phase makes the feature vector dimension high.An improved feature extraction algorithm based on Gabor wavelet transform is proposed,which enhances the feature relationship of each pixel in the local neighborhood,and integrates the amplitude and phase features.Finally,simulation experiments show that the improved algorithm can extract low dimension feature vectors,improve the representation ability of feature vectors,and ensure the high robustness of the eigenvector to the illumination change.(4)Aiming at the search strategy of the existing feature selection algorithm is more complex.A feature selection algorithm based on Monte Carlo Tree Search(MCTS)is proposed,which translates feature selection problem into sequential decision problem,and uses Monte Carlo tree search strategy for feature subset selection and choose the Relief-F algorithm with small computational complexity as the evaluation function of the search process.The simulation experiment shows that the algorithm has high efficiency and can extract better feature subsets with low dimension and separable.