Deep Low Resolution Face Recognition Using Identity Constraints

The technology of face recognition has been widely used in finance,medical care,security and other fields.From the beginning,it was only used for a single circumstance,to now it needs to deal with lighting,posture,covering and other complex situations.Although the existing face recognition algorithms have been able to achieve good performance under the constrained conditions,in many actual unconstrained environments,due to the distance between the camera and the photographed face and height changes,the resolution of the collected face image is low,thus resulting in a significant decrease in the actual recognition performance.Low quality and low resolution of face recognition are important issues in current face recognition field.Deep learning technology shows good potential in the current image classification and recognition tasks and super-resolution technology can effectively improve the image resolution,which is beneficial to the subsequent image classification and recognition tasks.For this reason,this paper is devoted to the research of deep learning on low resolution face recognition method applied image super resolution technology.In particular,some preliminary explorations are made on how to make full use of the identity prior information of face to effectively improve the depth feature extraction of low resolution face image to improve the recognition performance.The specific work of this paper is as follows:1)A low resolution deep network for face recognition based on high resolution image assisted feature enhancement learning is proposed.The main idea is that in the phase of deep learning,on the one hand,use high resolution sample image to constrain super-resolution reconstruction of low resolution image.On the other hand,use cognitive features and status of auxiliary information of high resolution image to extract the depth of super-resolution image feature,improving the capacity of the expression of the depth of the low resolution image characteristics,so as to improve the recognition performance.Specifically,the designed deep network consists of two channels,one of which is a low-resolution face recognition sub-network,which performs the super-resolution reconstruction,depth feature extraction and recognition of the input low-resolution faces sequentially.The other channel is the high-resolution face recognition sub-network,which sequentially extracts and recognizes the deep features of the input high-resolution faces,and assists the learning of the low-resolution face recognition network in the network training stage.The network training combines the pixel loss,multi-level feature loss and identity loss of the super-resolution face,and effectively improves the feature representation and discrimination ability of the low-resolution face,resulting in a significant improvement in the recognition performance.Experimental results verify the feasibility and effectiveness of the proposed method,and the recognition performance is improved significantly even in the case of very low resolution.2)A low resolution face recognition method based on progressive super-resolution reconstruction is proposed.The main motivation is to use the multi-scale image priori to guide the progressive super-resolution reconstruction of the face image,and use the multi-scale image identity information to constrain the depth feature enhancement extraction of the low-resolution image,so as to improve the overall performance of the low-resolution face recognition network.Firstly,a multi-level cascade super resolution network is used to enhance the progressive resolution of low resolution faces,and then face features of different resolutions are extracted simultaneously for face recognition.The network training also refers to the pixel loss and identity loss of the label image.A large number of experimental results verify the effectiveness of the proposed method.