Studies Of Low-resolution Face Recognition Technology Under Uncontrolled Conditions

Face recognition technology has been extensively researched and practically applied because of its advantages such as simplicity,convenience,safety,reliability,and userfriendliness.The existing face recognition algorithms mainly perform matching and recognition for high-quality face images under specific conditions and have achieved great success in terms of recognition accuracy.However,with the development of information technology and based on the needs of public safety,the monitoring system is gradually improved and the monitoring network has almost achieved full coverage.Due to factors such as unsatisfactory shooting scenes and random behavior of the person being photographed,the quality of the images captured by these surveillance probes is low,which leads to the existing face recognition algorithms have poor performance on such low-quality face images.Previous studies have shown that the main difficulty of lowquality face recognition technology is the low resolution of the image,which leads to the fact that existing face recognition algorithms cannot extract facial feature information with sufficient identity discrimination.Therefore,based on the theory of deep learning,this dissertation proposed the FSRCNN-sub algorithm and FSRCNN-subs algorithm respectively for low-resolution face images super-resolution by improving the FSRCNN algorithm,that is replacing the deconvolution layer in the FSRCNN network structure with a sub-pixel convolution layer and using the Adam algorithm to replace the SGD algorithm.Then we use the Face Net network and the VGG-Face network to extract facial features from the reconstructed highresolution face images.And finally,we complete the face matching recognition by calculating and comparing the cosine distance between the facial features.Through the calculation of the PSNR value of the reconstructed high-resolution image and the face verification experiment on the LFW dataset sampled at different magnifications,the results showed that the performance of the proposed algorithm is better than the original FSRCNN algorithm.In addition,in order to observe the facial features finally extracted by the network and explain the decision basis for the network to give the final classification and recognition results,this dissertation also used the guided back propagation,Grad-CAM and Guided Grad-CAM algorithms to visualize the VGG-Face face recognition network.