Research On Low-resolution Face Reconigtion Based On Deep Learning

With the rapid growth of modern society's demand for face recognition tasks,video surveillance systems are increasingly being used in security and forensics fields.Therefore,in some complex and unconstrained scenes,such as urban streets with very crowded people,it is usually necessary to perform face recognition on the faces of some pedestrians captured by the camera.But in most cases,these captured faces have other factors that interfere with the face recognition task,such as changes in light intensity and different angles where the camera is located,which will lead to large differences in the features extracted by the face recognition model.These factors will undoubtedly affect the accuracy of face recognition tasks.In addition,due to the excessive distance between the object of interest and the camera,the insufficient resolution of face photos is also very common.In face recognition tasks,it is necessary to match features extracted from low-resolution face images and high resolution face images.Since these low-resolution faces have less information and more noise,how to match images of different resolutions will be a practical but challenging task.For low-resolution face recognition tasks,we started from two aspects,improved the traditional face recognition methods,and achieved good results.The research content of this article is as follows:(1)Aiming at the problem of low-resolution face recognition based on embedding,a low-resolution face recognition network based on progressive partial coupling strategy is proposed.For the design of the network model,we adopted a progressive coupling strategy to explore the impact of coupling convolutional layers of different depths on the experimental results.In the process of forward propagation,since the generated features of the HR branch network and LR branch network become more and more similar as the network depth increases,so we combine different blocks in the network from back to front to find the optimal coupling choice.Using this strategy,we can further improve the recognition performance of Sphere Face on low-resolution face images.Compared with the state-of-the-art method,the performance of our proposed method on the LFW dataset with low-resolution face images of 14×12 and 7×6 pixels as input increases by 7.34%and 21.30%,respectively.(2)Aiming at the problem of low-resolution face recognition based on super-resolution reconstruction,a low-resolution face recognition method based on spatial attention is proposed.Our solution is to introduce a spatial attention mechanism to make the face image reconstructed by the super-resolution sub-network highlight the parts that are beneficial to face recognition.To reduce the influence of domain shift on experimental results,we have designed an adaptive module to learn the mapping from SR features to HR features.Compared with SRLRFR which is based on super-resolution reconstruction,our scheme can achieve better results.On the LFW data set,the performance of 28×24,14×12,and 7×6low-resolution face images as input increased by 0.35%,3.48%,and 2.73%,respectively.