Research On Face Super Resolution Reconstruction Technology Based On Deep Learning

For the past few years,computer vision has made great progress against this background of great development of deep learning.Face super resolution technology is a technology to enhance image resolution in computer vision.The practical significance of this technique is to reconstruct low resolution face image to obtain high resolution face image.Face image is widely used in advanced face image fields such as face detection,face recognition and face key point detection.Because of face image is easy to lose details in the process of collection and transmission,face hallucination technology is particularly important for advanced face image processing.Although the current face hallucination algorithm has experienced a fundamental leap from traditional methods to the deep learning methods,and constantly updated and iterated new algorithms,but in advanced face image tasks,due to the continuous improvement of the quality of the input image clarity requirements,as well as the improvement of the complexity of the collected image scene,face image resolution reconstruction requirements are constantly improved.In this context,it is of great significance to develop a method to improve the accuracy of face super-resolution algorithm.In order to solve the problem that the contour information of face image is difficult to recover,the method of combining super-resolution network and prior information network is selected to guide the image re-construction.In order to solve the problem of high smoothness of texture details in key parts,a feedback loop training method is proposed.The network also introduces an adaptive deviation modulator to solve the problem of pixel deviation in the reconstruction process.Specific research contents are as follows:(1)An improved algorithm based on attention-directed cascade face super resolution is proposed.This method uses FSR-Net face superresolution algorithm as the baseline algorithm.Firstly,to solve the problem that FSR-Net network is not enough to extract important features from feature blocks and global information loss exists in the process of feature extraction,a Parallel attention mechanism(PAM)guided lightweight residual structure is constructed to extract feature information.It realizes the efficient utilization of feature information.Secondly,texture loss is introduced into the loss function to make the reconstructed face image close to the real image at the edge of the face.In Helen test set,the PSNR and SSIM indexes of the improved face super resolution network are increased by 0.58 d B and 0.0309 respectively.(2)An improved face reallocation method based on DIC face super resolution algorithm based on adaptive bias modulation attention is proposed.First,to solve the problem that Batch Normalization(BN)is abandoned in the deep face super-resolution module,which leads to the large pixel deviation of the reconstructed image,an adaptive deviation modulator is designed.The structure can adjust the numerical change,so as to offset the error caused by the introduction of BN layer,so that the global network can achieve faster training and better generalization.Secondly,in order to solve the problem of the imprecision of the facial prior alignment module,the combination of attention mechanism(PAM)and face re-alignment is used.The attention mechanism includes multilayer perceptron sensitive to position information.In the rough super resolution module training,PAM initially corrects the position of facial organs in LR images.Face re-alignment is a compensation measure to make up for the inaccuracy of the key points.The key points of face can be accurately located by two face alignments.The global network adopts feedback loop training mode to realize feature reuse.The PSNR index of the algorithm is improved by 1.1d B compared with FSR-Net and 0.15 d B compared with DIC.(3)An improved face super resolution reconstruction method based on dense residual connection and cross-scale fusion is proposed.Firstly,to solve the problem of texture smoothing caused by the loss of high frequency feature information of the model,this paper takes the improved RCAN algorithm as the backbone network of the super-resolution module,and adopts the dense connected residual channel attention structure inside the super-resolution structure to fully capture the network feature information of each layer.Secondly,to solve the problem that the superresolution reconstruction of face key point location is not obvious,an improved key point detection branch structure is introduced.The main body of this module is a combination of lightweight HR-Net algorithm and residual self-attention.In this method,the features of backbone network feature characterization always maintain high-resolution.In order to avoid the problem of inaccurate positioning at the key points with dense pixels caused by the use of subsampled feature maps,residual self-attention is used to solve the problem of information sharing between different subsampled feature maps.Compared with DIC,PSNR and SSIM improved by0.25 d B and 0.0128 respectively.