Study On Image Super-resolution Reconstruction Model Based On Neural Network

Image super-resolution reconstruction aims at reconstructing corresponding high-resolution images from low-resolution images.It is widely used in security monitoring,Internet information transmission and medical imaging.Traditional image super-resolution reconstruction methods,such as sparse coding and manifold learning,are limited by limited manual feature extraction methods,and new ideas need to be sought to improve image reconstruction quality.The deep learning method emerged in recent years has made up for the shortcomings of traditional methods.It can automatically learn features from data to improve the reconstruction results,and initially shows great potential.Due to the ignorance of good quality feature in other layers,methods based on single-layer feature can not get top reconstruction results.And methods based on multi-layer features need to explore an effective way to reduce the proportion of poor quality features.Therefore,we needs to explore a multi-layer feature based reconstruction model,which can reduce the proportion of poor features when reconstructing.The main contributions of this dissertation are as follows:1)We analyze the limitations of some existing models,both single-layer feature and multi-layer features based models are included,e.g.VDSR and DRCN.Our evaluation experiment conducted on VDSR shows that using multi-layer features is helpful to reconstruction.Besides,we contrast DRCN with VDSR and find out that not reducing the proportion of poor features is responsible for the limitation of performance of DRCN.2)We design BySEDenseSR,which is based on by-pass convolution and feature weighting,to reduce the bad influence of poor features.On the one hand,we design a network based on squeeze excitation block.It dynamically compute the weights of different features,thus can set small weights on poor features.On the other hand,using by-passing convolution,BySEDenseSR can reduce forwarding noises,improve the stability of middle-layer features.Experiments show that compared to DRCN,BySEDenseSR improves its PSNR by 0.17dB,0.25dB and 0.12dB respectively on Set5,Set14 and BSD 100,at magnification 4.3)Improve upon BySEDenseSR,we utilize step-by-step reconstruction technique to cut down its time and space complexity,propose a new model called BySEDenseSR+,it also gives better reconstruction results.In BySEDenSR,the input image need conduct interpolation first,resulting in its time and space complexity leap with its input size increasing.We introduce in Laplacian pyramid structure to eliminate this situation.Experiments show that compared to BySEDenseSR,BySEDenseSR+gets 5.5 times speed up,and improves its PSNR by 0.05dB,0.06dB and 0.01dB respectively on Set5,Set14 and BSD100,at magnification 4.