Research On Face Recognition Algorithm Based On Deep Neural Network

With the development of society and the advancement of science and technology,the core problem in the field of social security is no longer the comprehensive monitoring coverage,but the rapid and effective automatic identification under the comprehensive monitoring coverage.Face recognition technology is proactive and non-invasive,and is an important means of rapid automated identification and an important research direction for pattern recognition.The face recognition algorithm based on Deep Neural Network can automatically extract facial image features and complete classification.Its excellent recognition effect makes it widely used in existing face recognition systems.This paper focuses on two common Deep Neural Networks: Deep Believe Network and Convolution Neural Network.The main work of this paper is as follows:Firstly,for the shortcomings of Deep Belief Network,the learning effect of face image samples is easily affected by external conditions such as illumination and occlusion.A new face image preprocessing algorithm is designed,including illumination processing,occlusion processing and dimensionality reduction processing.The illumination method changed the threshold segmentation method in the Self Quotient Image(SQI).Experiments show that the improved SQI is more robust to illumination.The global gray density function is added to the pressure function in the Level Set(SBGFRLS)during occlusion processing,the experimental results show that the effect is better when the face image with small gray scale change of the whole image is occluded.In the dimension reduction process,the block information entropy is used in the Two-Dimensional Principal Component Analysis(2DPCA)to determine the contribution of the sub-block feature vector to the projection matrix.It can automatically and effectively eliminate the high-dimensional redundancy attribute of the face image while more characterizing the face image information.After the face image is preprocessed,the preliminary features of the face can be obtained to eliminate illumination,occlusion and high dimensional redundancy.Then,using the initial facial features as the input data of the Deep Belief Network,the layered Restricted Boltzmann Machine(RBM)is used to continue to learn the essential category information implied in the preliminary features.At the top level,Softmax is used to classify and obtain the recognition result.Compared with other algorithms under the same conditions of classification method and hidden layer,the results show that the proposed algorithm has higher recognition rate.In addition,for the disadvantage that the Convolution Neural Network can only extract the whole face information with the same granularity,an improved face recognition algorithm based on Convolution Neural Network is proposed:First divide the original image and calculate the information entropy of each sub block.First,the original image is divided into blocks,and the information entropy of each sub-block is calculated.Then,preliminary feature extraction is performed by allocating convolution kernels of different granularities according to the amount of information contained in the image sub-blocks.Finally,the feature information of all sub-blocks is merged and the classification is completed.Using the improved face recognition algorithm based on Convolution Neural Network on each face database compared with the same type of algorithm.The experimental results show that the proposed algorithm has a high recognition rate and proves the feasibility of the proposed algorithm.