Research On Structure And Feature Metrics In CNN Face Recognition Model

In recent years,convolution neural network has been used in different areas of computer vision with its dramatic evolution.The rapid growth of computing resources and datasets makes it possible for convolution neural network to improve the accuracy of face recognition.With the increased requirements for performance,deep face network designing has become a key factor in face network performance evaluation.Existing networks mainly add multi-layer convolution layer to extract more efficient face feature vectors,which makes the network contain many parameters and ignore the importance of width.Simultaneously,the relationship between the extracted face feature vectors by deep face networks and statistical methods is not well analyzed.In this paper,an optimal method for networks design is proposed based on the analysis of existing deep face networks and used in designing a new balance network structure.In addition,an evaluation method is designed based on mutual information theory.(1)With the analysis of existing face networks such as DeepID,DeepID2,FaceNet,an optimal method in deep face network design is proposed based on the concept of depth and width.It can be described as: When the depth and width of the deep face network structure are close to each other,the recognition effect is improved significantly.The depth of the face network designed by this optimization method can constrain the network structure in a certain depth range and avoid the over-fitting in the training while obtaining the ideal recognition accuracy.(2)A new multi-level deep face recognition balance network is designed in this paper by using the proposed network design optimal method and ResNet.The performance of designed networks in different face datasets such as FERET,FRGC,LFW and YouTube Face shows that the proposed balance network works better with the help of optimal method.The depth of balance network is reduced and has high robustness.(3)A criterion based on mutual information is proposed.And then it is used to compare the characteristics of different face recognition features such as LBP,Gabor and deep learning face features.The results show that deep face features can maintain more useful face information when compared with statistical face features,which helps deep face network achieve higher accuracy.