Face Detection Algorithm Based On Deep Convolution Neural Network

Face detection is an important part in the field of pattern recognition and computer vision.It is also the research premise of human face such as face alignment,face recognition and facial expression recognition.Traditional face detection algorithm,usually using manual feature extraction algorithm for face feature extraction.In practice,the performance and robustness of the face detection algorithm based on manual design features need to be improved due to face occlusion,multi-posture,angle changes,light changes and facial expression changes.In view of the above problems,face detection based on convolution neural network has been widely studied.When traditional convolution neural network is used for face detection,the sliding window or Selective Search algorithm is usually used for region extraction.This kind of method has the disadvantages of complex algorithm,Time-consuming and other issues;at the same time,when the network depth reaches a certain level,the problem of gradient dissipation or gradient explosion occurs during the training,which leads to the degradation of network performance.Therefore,aiming at the problems of traditional convolution neural network in face detection,this paper studies the face detection algorithm to improve the performance and robustness of face detection.This article mainly completed the following work:(1)The current research status of face detection and deep learning is summarized through reading a large amount of relevant literature at home and abroad,and the problems existing in face detection are summarized.The research direction and main research contents of this paper are confirmed.(2)Firstly,the composition of convolutional neural network is introduced respectively from convolution layer,pooling layer and activation function,followed by the training method of convolution neural network,and then the stochastic gradient descent algorithm of convolution neural network is introduced.Finally,Focus on the Caffe depth learning framework.(3)Based on the classic AlexNet network,the image classification framework is designed and improved.In order to increase the ability of non-linear expression and reduce the number of parameters,the larger convolution kernel is replaced by two smaller convolution kernels.In order to solve the problem that the size of the data set is not uniform,Layer airspace pyramid pool processing.In this paper,the regional feature extraction algorithm in Faster R-CNN network is introduced.The region generation network firstly generates candidate blocks for batch images and then uses the convolutional neural network to optimize the location and size of the candidate boxes to obtain the classification probability,then generates the cost function,And then through the inverse parameter optimization,to improve the probability of region extraction.In this paper,we use FDDB and Wider face database to conduct experiments.The experimental results show that the proposed algorithm has an accuracy of 98% and an average detection time of 0.32 s.(4)Based on the theory of residual learning,construct a residual convolution neural network with a depth of 34 layers.In order to ensure that the network is easy to train and converge quickly,a random parameter initialization method is introduced in this paper.The method uses a mean of 0 and the standard deviation is based on the current layer The convolution kernel size and the number of convolution kernels are calculated.The initialization method can speed up the convergence of the network.Experiments show that in the FDDB,Wider face database,this method achieves 98.6% accuracy.