Research And Application Of Intelligent Face Recognition Methods

Face recognition is currently one of the most popular biotechnology.It has been widely used in security,access,justice and other fields.However,in the actual environment to address the deficiencies of existing face recognition algorithms,the specific research work here has the following aspects:(1)For deep belief networks during the facial feature extraction directly ignore the local characteristics of the problem,at the same time considering the traditional local ternary pattern of texture feature description is not enough,eventually to extract the local information is not detailed,so this paper proposes local ternary pattern based on the improved local mode and the combination of deep belief networks face recognition algorithm,using part of the traditional local ternary pattern of the surrounding pixels and the pixels in the middle of the difference comparing with a fixed threshold to change in order to compare two adjacent pixels,and then their difference comparing with calculated dynamic threshold,At the same time,the rotation invariance of local binary mode is introduced into the improved local ternary pattern to prevent the recognition rate from being reduced due to the rotation problem.The next step is to extract the feature of the improved local ternary pattern as the input of the depth confidence network,and then the parameter tuning of the depth confidence network is carried out to complete the recognition task.The effectiveness of the algorithm is verified by experiments in public face database.(2)Although the DenseNet network uses dense connections to enhance the transfer of features,the conversion of feature maps increases and the amount of redundant information generated continues to increase.It can't distinguish the importance of these features to the classification in the transfer process.If all the features are transferred without difference,it will lead to training errors.In response to this problem,the residual focusing mechanism is released to the DenseNet network and proposed A densely connected convolutional neural network model based on residual attention.The network can pay more attention to the features that are useful for classification and ignore the features that have little effect on classification,effectively reducing the redundancy of information.Take the improved network as the backbone extraction network of the FaceNet network.Function training requires a large sample set and is difficult to train.A loss function combining Center Loss and Softmax Loss is proposed to replace Triplet Loss as the final loss function,so that the reduced sample set can be used to obtain energy.Very good training effect.Finally,the improved FaceNet network is compared with other excellent networks,and a better recognition accuracy rate is obtained on the public data set.(3)Finally,the improved FaceNet model algorithm is applied to the actual scene,an online real-time face recognition system is designed,and the system is tested in terms of face pose,lighting,occlusion,etc.to verify the performance of the system.