Research On Multi-branch Fusion Single Image Super-resolution Reconstruction Method Based On Attention Mechanism

In order to solve the problem of poor quality of image restoration caused by insufficient feature extraction and the same weights given to low and high-frequency informtaion in super-resolution reconstruction,a single-image super-resolution reconstruction method based on attention mechanism and multi-branch fusion is proposed.First of all,in the feature extraction stage,three layers of dilated convolution with different expansion factors are used to effectively expand the receptive fields,which can extract image feature information better and avoid the problem of information loss caused by continuous use of dilated convolution.Secondly,a multi-branch fusion structure is constructed in the residual network to extract different feature information of images under different receptive fields and perform organic fusion,so as to strengthen the information flow in the whole network and extract the image features better.Finally,the attention mechanism is introduced to the residual network.By using the channel and spatial attention mechanism,the weights of image feature are redistributed to strengthen the weight of high-frequency information.Thus,the network pays more attention to the edge and texture information of the image,so as to improve the image reconstruction effect.By using the public data sets which are Set5,Set14,BSD100 and Urban100,the proposed method is compared with different methods including Bicubic,SRCNN,FSRCNN,VDSR,EDSR and RCAN.When the magnification is 2,3,4,the PSNR increased by 4.01 db,2.12 db,1.92 db,1.42 db,0.45 db and 0.30 db on average respectively,and the structural similarity increased by 0.087,0.039,0.034,0.024,0.007 and 0.004 on average respectively.Experimental results show that the proposed method can better restore the high-frequency details of the image to improve the image reconstruction quality,and achieve good results in objective quality evaluation and visual effect.This thesis has 46 figures,13 tables,and 55 references.