Research On Face Recognition Based On Convolution Neural Network

The characteristics of local connection,weight sharing,pooling operation and the powerful automatic extraction feature effectively reduce the complexity of the network structure,and obtain good robustness for image scaling,translation and distortion invariance.So the convolutional neural networks can effectively improve the recognition performance if it is applied to face recognition.This paper studies the architecture of convolutional neural network and corresponding solutions are proposed for different problems.In order to reduce the loss of information in the process of transmission derived from multiple pooling operation in the traditional convolutional neural network,The shortcut connection model is proposed that the shallow and deep feature are integrated so that the feature loss can be reduced.At the same time,in order to obtain more distinct feature,our algorithm adds a penalty to the original softmax loss function so that the distance can be minimized between the learned class feature.As a result,the loss of the detail feature are reduced and the discrimination of depth feature are increased.To solve the problem thatthe multi-scale information of the image cannot be made full of use in the traditional convolutional neural network,the multi-scale residual network model is proposed.the feature characterization is improved through convolution fusion of multiscale and cross channel.At the same time,in order to reduce the learned parameters of the deep network architecture and make it easy to be optimized.The feature of the previous layer is concatenate with the feature of the next layer across the multi-scale module layer directly and then the large convolution kernel is decomposed into a small asymmetric convolution kernel which reduces network parameter calculation.In order to validate the performance of the model,the graphical interface is implemented with the asymmetric little convolution model.The experimental results show that the algorithm has good robustness to the scale change,occlusion,complexity background and so on.