A Study On Occluded Face Recognition Based On Mix-Squeezenet

With the application of convolutional neural networks in face recognition technology in recent years,people have designed more and more layers of convolutional neural networks to improve recognition accuracy.Such models are generally difficult to transplant to some small devices such as mobile devices.Device side.In addition,face occlusion recognition is a difficulty in face recognition technology,mainly because the presence of occlusion on the face will cause the loss of local features in the process of feature extraction,and it is easy to be confused with the local unoccluded features of the face Problems such as masks and scarves obscure the loss of features extracted by the neural network in the lower half of the face.Combining the above problems,there is a lot of room for development in the research of lightweight face occlusion recognition.This article has mainly done two aspects of work.First of all,in order to be able to run on a small device,we proposed a lightweight convolution fusion model based on squeezenet which is mix-squeezenet.We discussed the way to improve the performance of feature extraction and analyzed the core structure of the basic model squeezenet.,We use the small convolution kernel of the receptive field to replace the large convolution kernel in squeezenet,so that the convolutional neural network has stronger nonlinear performance,and the pooling layer is used to characterize the subsequent output fire module of equal size Fusion makes the extracted features richer in information,and prevents the problem of small convolution kernels that are easy to overfit on complex face training.After the above improvements can reduce the training parameters,the model can more easily converge and improve the performance of lightweight network face recognition.Then for the poor recognition effect after partial occlusion under the face,we use a certain proportion of occlusion training on the image data set to enhance the robustness of the network model for occlusion recognition.In addition,we need to reduce occlusion during image training The influence of the region on feature extraction,so the cut-off 2D Gaussian kernel function is used to process the blocked image,and the shape of the specific cut-off 2D Gaussian kernel function is designed according to the impact of occlusion on the extracted features,and the function is directly applied to After training the FeatureMap,setting different weights for different regions can effectively reduce the impact of the partial occlusion of the face on the overall feature extraction and try to keep the features of the unoccluded upper part of the face area.Finally,we train on the CASIA-WebFace training set.For the mix-squeezenet,the accuracy of the model is experimentally verified in the LFW test set with different occlusion ratios,which has better results than other models.