Research And Implementation Of Image Super-resolution Based On Blur Kernel Estimation And Attention Mechanism

Super-resolution task aims to restore a given original low-resolution image to a highresolution image.With the development of deep learning,image super-resolution technology has important application value and play a vital role in many fields such as face recognition,medical image processing,target detection,surveillance equipment and satellite images.At present,the research methods of super-resolution mainly include traditional algorithms and deep learning algorithms based on neural networks.Although deep learning methods based on neural networks have achieved a great improvement in reconstruction quality and reconstruction efficiency in the field of Single Image Super Resolution(SISR)compared with traditional algorithms.Butthe learning ability of deep learning methods is mainly determined by the quality of the training data,and the training data of the existing super-resolution network model is artificially synthesized,that is,it is assumed that the blur kernel during the downsampling period is predefined or known(for example,Bicubic downsampling),but the blur kernel involved in the real scene of the actual application is very complicated and unknown.The difference of data distribution between the dataset used in the real world and the dataset used for training will cause a sever drop on Super-resolution performance.Therefore,it is still a challenging problem how to only use the existing low-resolution images obtained in real scenes to achieve superior super-resolution effects in real scenes.In order to solve the above challenges,this article mainly carried out three aspects of innovative work.(1)Blur kernel as a prior information is embedded to generate low-resolution images to improve the generalization ability of super-resolution networks.Blur kernels have prior information to guide image reconstruction,so the blur kernel estimation method is used to obtain the blur kernels,and the blur kernels can be embedded in the generative countermeasure network to obtain complex and degraded low-resolution image datasets.Use this data set to train the super-resolution network can improve the generalization ability of the model.(2)The prior information of the facial landmark is embedded to maintain the information of the geometric structure of the face.Taking into account the instability of the generative confrontation network,facial landmark is embedded in the network to maintain the geometric structure information and alleviate the face deformation problem.(3)We employ attention mechanism to increase the visiual quality of the results.In order to improve the reconstruction quality of the image texture and edge details of the super-resolution network,the attention mechanism is used to pay attention to the detailed texture information that is difficult to recover during the reconstruction process,and to suppress the interference features.In addition,in order to alleviate the problem of gradient disappearance and speed up training,the residual structure and attention mechanism are introduced to reconstruct images.The network of this paper has been qualitatively and quantitatively tested on LS3 DW and LR-testset.The results of our experiments show that our complete model has a good generalization ability for real world datasets.In addition,on the real data set of LRtestset,ablation experiments were carried out on the three components of the blur kernel,facial landmark and the channel attention mechanism.The experimental results show that embedding the prior information of the blur kernel and facial landmark into model is effective and plays an active role in generating LR datasets.