Face Recognition Based On Deep Learning

Object recognition is one of the important field in artificial intelligence,which is mainly to make the computer identify different objects more efficiently.Face recognition is an important embodmient of human-computer interaction,which has been widely used in many fields,such as access control system,video surveillance,and identity verification.However,face images easily vary with expressions,ages,as well as poses of people and environment of imaging.In addition,environment of imaging also includes illumination,shade and other factors.Therefore,face images are rich and easy to change.The variability of face images greatly increases the difficulty of face recognition,which enhances the requirement to the algorithm that extracts the face features.Face recognition requries more accurate recognition rate,as well as less training time,which increases the difficulty of face recognition to some extent and makes it more challenging.This dissertation makes a deep study of face recognition,and an efficient algorithm of face recognition by combining deep Convolutional Neural Network(CNN)and Support Vector Machine(SVM)is obtained.In this algorithm,a deep Convolutional Neural Network is designed as the features extractor.The SVM is designed to be the classifier with the output of CNN as the input and commplish s the final face recognition.For CNN,a network with nine layers is designed to extract facial features automatically.In addition,a SVM with gauss kernel function is deaigned to classify facial features.This dissertation optimizes the network with some optimization techniques to improve the overall performance of the network in extracting features and the training speed.The main techniques include network overfitting prevention and the cross entropy instead of traditional quadratic cost function is used to accelerate feature extract of network.In addition,using lots of auxilizry data to pre-train CNN to improve the generalization ability and accelerate convergence of the network.According to the features extracted by CNN,face identity is judged by the voting mechanism after the multi classification of the Support Vector Machine is achieved by one to one method.In the image processing,horizontal inversion and rotation are used to expand the sample data to improve the learning ability of network.We find the effect of illumination to face recognition when comparing the recognition results among different training datasets and testing datasets,and then the illumination compensation method is utlized to process the images where features are not easily to extract.This dissertation takes advantage of the part of images from Casia-Webfaces database as the auxiliary data to pre-train CNN.The FERET database with face images in different pose,illumination,and expression is used to test the combination syetem of CNN and SVM.The final experiment shows the alglorithm has a relatively good recognition result in a short training time.The algorithm in this dissertation has superiority in recognition result and training time when comparing with the existing face recognition method.