Low-resolution Face Recognition Based On Deep Learning

Low-resolution face recognition has always been a difficult problem in the face recognition field.Its recognition accuracy is affected by many external factors,such as light intensity,face posture,and human face expression.In the case of one-to-one face verification,low resolution The rate recognition algorithm has been well implemented,but it has been difficult to recognize the identity of multiple faces.Therefore,this paper designs a low-resolution face recognition system based on this to target low-resolution face recognition.face-Multi identification,the main research content is as follows:First,a low-resolution human face detection network is designed to improve the detection accuracy of low-resolution human faces.The face detection network adopts the concept of joint-level,continuously screens and optimizes the detection results of the human face,and ultimately obtains relative The accurate face detection area enables effective extraction of face images in face-multi and low-resolution face environments,ensuring the accuracy of the first step.Secondly,the network design of the super-resolution reconstruction module is completed,and the super-resolution reconstruction methods under different networks are compared.The advantages of super-resolution reconstruction of low-resolution face images based on the deep residual network are determined,and multiple resolutions can be performed.For super-resolution reconstruction of low-resolution face images of various levels,the residual network of different depths is compared with the data from the experimental results to determine the network depth of the final experiment.In order to reflect the effect of super-resolution reconstruction,the reconstruction results of the interpolation-based super-resolution reconstruction method were compared and the experimental results were analyzed.Once again,a scheme is taken to train the face feature extraction network with each set of three face data as a set.The purpose is to improve the discrimination of the extracted face feature vectors,thereby improving the accuracy of face recognition,and then at the experimental level.Verify the feasibility of the method.Then,the design and implementation of neural network softmax classification and K-nearest neighbor classification are completed.The similarities and differences of the two classifiers in the algorithm are analyzed.When using the softmax classification method,due to its network The degree of freedom of design is higher.Therefore,softmax classification modules of different network structures are designed and compared with the recognition results of K-nearest neighbor classification method under different samples.Finally,the K-nearest neighbor classification method is used to determine the low sample size.Resolution face recognition.Finally,the experimental verification of the low resolution face recognition system is completed.In order to embody the recognition effect of different levels of low-resolution face and the effect of super-resolution reconstruction on low-resolution face recognition,the low-resolution face recognition accuracy of different pixels is compared,and the import and non-import super-resolution are compared.Reconstruct the recognition effect of the network and analyze the results of the experiment.