Research On Dynamic Face Recognition In Complex Environment Based On Deep Learning

Face image recognition has always been an important component of computer vision and image recognition.It is also the most widely used application for deep learning and big data.As a new neural network method with forward feedback,convolutional neural network is combined with deep learning technology.Compared with the traditional face recognition algorithm,which used artificially designed face feature extraction algorithm.Due to the methods of local connection and weight sharing,the convolutional neural network has better robustness and accuracy when face images in complex environnents have fuzzy,angular and illumination variation situations.This makes the convolution neural network widely used in the field of image recognition.In this paper,the detection method of face recognition is designed according to the structure of convolutional neural network,the pooling methods,over-fitting problem and face location are improved by updating the algorithm.This article mainly completed the following work:Firstly,the research status of face detection and deep learning is understood through reading domestic and foreign literature,the existing problems of face detection are summarized,and the research direction and content of this paper is determined.Secondly,the thesis summarizes the theoretical knowledge,deep learning framework and optimization algorithm of convolutional neural network.Firstly,the composition of convolutional neural network is introduced from convolution layer,pooling layer and activation function,then the training method of convolutional neural network is explained,then the stochastic gradient descent algorithm of convolutional neural network is introduced,and finally the TensorFlow deep learning framework is emphatically analyzed.Then,aiming at the disadvantages of single traditional pooling method and difficulty in extracting effective features in image recognition,an improved method is proposed for optimal pooling effect on pooling layer of convolutional neural network.Based on the optimal search theory,this paper studies how to maximize the probability of the detected objective function under the constraint conditions,and finds out the probability that should be assigned to the pooled results of each child element under different conditions,so that each child element can be allocated with the optimal probability when pooled,so that the convolution effect can be better.The pooled method of optimal search is compared with the pooled method commonly used in current applications.Then,considering the over-fitting problem of convolutional neural network in deep learning,some methods such as Weight decay and Dropout are proposed to solve the over-fitting problem.The weight value is corrected within a certain range by the addition of regularization.At the same time,the convolutional neural network structure was added to berserk to randomly close or ignore nodes of certain layers during model training,and the size of learning rate was adjusted according to the accuracy rate in the iteration process,and a comparative experiment was conducted.Finally,a filter algorithm based on GLMB(generalized labeled multi-Bernoulli)was proposed to solve the problem of difficulty in locating a multi-target face in a complex environment,the accuracy and robustness of face localization are improved.The experimental results show that the optimized pooling method is easier to extract effective features and has higher accuracy.The methods such as Weight decay and Dropout can effectively solve the over-fitting problem in the process of model training,it also reduced training time.The face location filter algorithm improvement based on GLMB(Generalized Labeled Multi-Bermoulli)can accurately locate multiple face targets in a variety of complex environments.