Research On Face Recognition Algorithm Based On Deep Convolutional Neural Network

Face recognition technology is a technique that uses computers to extract features and express data of faces in images.It classifies and recognizes identities according to extracted feature information.In a relatively ideal environment,face recognition has reached a high degree of precision,but in practical applications,uncontrollable factors such as illumination,posture and expression greatly restrict the recognition effect.Therefore,how to express facial features in complex environments through neural networks has become the focus of face recognition research.Deep Convolutional Neural Network is a kind of Deep Learning technology that automatically learns facial features without manual intervention.It is also one of the face recognition algorithms commonly used by scholars at home and abroad.The key point of Deep Convolutional Neural Network is to construct a reasonable and effective neural network model,and select the appropriate data set to train the network,then we obtain a good classification effect.In this paper,the network performance is improved by optimizing the structure of Deep Convolutional Neural Network,and we delved into feature extraction on the profile face.The main work done in this paper is as follows:(1)The research status of face recognition technology at home and abroad is investigated.The Adaboost face detection algorithm and Constrained Local Neural Fields feature point location algorithm are deeply researched and verified.The experimental results show that Constrained Local Neural Fields algorithm can accurately locate face feature points in actual scenes;(2)The network structure is analyzed by the effective receptive field theory.we use the global convolutional layer instead of the global average pooling layer to improve the structure of the 34-layer Residual Network.The network is trained and tested through the Casia-WebFace and LFW database.The test results show that the improved network model accuracy rate reached 98.3%;(3)This paper analyzes the problem of the profile face recognition on the improved network.By splicing the Deep Residual EquivAriant Mapping in the network,the equivalent mapping between the profile face and the frontal face is performed in the depth space,which improves the accuracy of the network model in identifying the profile face;(4)Three different network structure models are designed and verified by comparative analysis.We choose the 28-layer network stitching Deep Residual EquivAriant Mapping model as the network model for the actual scene test.The test results in the actual scene show that the network model can effectively reduce the Euclidean distance between the profile face and the frontal face,and enhance the generalization of the algorithm to the face recognition.