Research On Lightweight Face Recognition DCNN

At present,many face recognition systems use deep convolution neural networks(DCNN)technology.Tested on the LFW dataset,the average accuracy of these systems has reached more than 99%.However,there are not many practical face recognition products or services.There are many reasons.One of the main reasons is that the DCNN model evaluated on the LFW dataset is large and requires a lot of computing resources,but mobile and embedded devices cannot afford it.Tosolve this problem,this paper focuses on the design of lightweight face recognition DCNN.The main work is as follows:In terms of network structure design,this paper uses depth-separated convolution instead of standard convolution and global averaged pooling layer instead of full-connected layer,which greatly reduces the number of parameters and computation of the network model and reduces the size of the model.For the problem that the softmax classifier is inflexible in practical applications,the idea that the DCNN directly learns the mapping of the original picture to the Euclidean distance space is proposed,and the similarity between faces is measured by the distance of the Euclidean space.In order to improve the training speed of the network,the performances of various activation functions and optimization algorithms were compared.It was found that using the ELU activation function and the Adam optimization algorithm can effectively speed up the training.In order to cooperate with the small batch training network,Each batch of training is randomly divided into mini-batch.The network also uses batch normalization to speed up network training and avoid overfitting.In order to further reduce the model size,the use of equal-area quantized joint Huffman coding for the network-trainable parameters reduces the storage space occupied by the parameters.During the training,the training set was expanded by using a random cutting method to enhance the generalization ability of the model.In addition,in the case of using triplet loss to learn the compact representation of facial images in the European space,this paper experimentally proves that the use of grayscale image training is better than the model generalization effect obtained by training with color images.And it has drastically reduced the amount of calculations and accelerated network training.Finally,there are correlations between the elements of the face feature vector output for the trained network,It is pointed out that the direct use of Euclidean distance to measure the similarity of two face feature vectors may be problematic.This paper proposes the use of Mahalanobis distance to measure the similarity of two face feature vectors.