A Sparse Residual Deep Network And Its Application In Face Recognition

Artificial neural network is based on the basic principle of neural network in biology.Based on the network topology,it simulates a mathematical model of the brain nervous system information processing mechanism to the outside world.Although the research on artificial neural networks experienced a rapid development at the end of the last century,it was at a low point due to various conditions at the time.With the rapid development of computer technology,especially the rapid increase of GPU performance in hardware,the new artificial intelligence model represented by deep learning has excellent performance in many fields and has rapidly become a research hotspot.In the field of driverless,image recognition,natural language processing,speech recognition,etc.,deep learning technology has a very wide range of applications.Convolutional neural network is a typical representative of deep learning technology,and it has remarkable performance in image recognition,target detection and other fields.Convolutional neural network is a special deep learning model,which includes convolutional and pooling layer.Because it adopts a series of operations such as local connection,weight sharing and subsampling,it greatly reduces the weight parameters that need to be trained.Therefore,the convolutional neural network can make the network highly robust in the case of many layers,and suppresses the problem of overfitting effectively.And convolutional neural network has many advantages such as translating and scaling invariance.In many convolutional neural network models,the residual network(ResNet)is representative.The residual network uses skip-connection between the input layer and the output layer,effectively suppressing the problem of the gradient disappearing.Because of the role of skip-connections,gradient information can be more efficiently transmitted from the back network layer to the shallow network.However,the skip-connection makes the entire network overemphasize the initial network input information,reducing the diversity of the extracted information.For this problem,this paper proposes a new residual network--Sparse Residual Network(SRN),which relies on a new connection method – sparse-connection.It makes more input layers connect to output layer and enriches the feature extraction of the network,so sparse-connection can effectively solve the above problems.Experiments have shown that SRN has achieved significant results and performed better than ResNet in both CIFAR-10 and CIFAR-100 datasets.Face recognition has always been a research hotspot in computer vision.This paper applies Sparse Residual Network to face recognition as an extractor of facial features,and build a face recognition system,including three modules: face detection,face preprocessing,feature extraction and verification.The system is trained in the CASIA-Webface dataset and tested on the LFW dataset.Experimental results show that SRN-based face recognition system has good recognition effect and real-time performance...