Research On Local Occlusion Face Recognition Method Based On Deep Learning

As a popular research direction of biomarker recognition technology,face recognition can be widely used in ID authentication?access control?Human—Machine Interaction?multimedia management monitoring,et al.Recently,although most face recognition technologies have achieved a better recognition rate under restricted conditions,when the external conditions change significantly,the face occlusion will result in incomplete face image features so that the face image cannot be compared with the face information in the library.The difficulty of face recognition caused by occlusion is mainly reflected in feature loss,alignment error and local aliasing.In this paper,a new deep learning model is proposed to recognize occluded faces.The main research works are:(1)A new data augmentation method is proposed,and a fine-tuning convolutional neural network is used to generate the occluded face image.This method mainly adopts the technique of occlusion sensitivity experiment to recognize CNN and extract the most distinctive face region.In this method,the recognized face area is firstly covered to force CNN to extract discriminant features from the unconstrained face area,in order to reduce the dependence of the model on the recognized face area in the process of training.The experimental results show that the CNN model trained by this method can significantly improve the performance on the real occluded face image showing in the AR face database.(2)We proposed the regional attention module to locate the unconstrained face image.The module makes the deep learning network focus on processing the unconstrained part of the face image.Secondly,an improved triple loss function is proposed to construct a network model for occluded face recognition,which integrates the following factors:1)That narrows the distance of the non-occluded face and the occluded face of the same people and expands the distance among different people of occluded faces and the distance between different people of occluded and non-occluded faces larger.2)That combines the positional attention mechanism with the channel attention mechanism,and the model can focus on processing non-occluded face parts.Simultaneously,it exploits the features of the non-occluded face parts efficiently and embeds the facial features well.The experimental results show that the dual attention mechanism can improve the recognition rate of the deep learning network under occlusion,and in the case of low occlusion,the channel attention mechanism contributes more to the network recognition rate than the position attention mechanism.