Research On Multi-scale Super-resolution Reconstruction Algorithm Based On Residual Back Projection Network

With the continuous development of deep learning,more and more results have appeared in the field of super-resolution reconstruction.Although the super-resolution reconstruction algorithms can achieve relatively good reconstruction results,there are still some shortcomings that need to be improved.In order to solve the problems existing in the existing algorithm framework,this article has carried out in-depth research and put forward some feasible solutions.The main research contents of this article are:Aiming at the problem that mainstream algorithms only perform feature extraction on a single resolution space,which cannot fully reflect the non-linear mapping relationship between low-resolution images and high-resolution images,this paper designs a super-resolution based on dense residual back projection.Rate reconstruction network framework.It uses an iterative up-sampling sampling network,which optimizes reconstruction errors by alternately using up-sampling and down-sampling as an efficient iterative process,and digs deeper into the direct interdependence of lowresolution and high-resolution images.At the same time,local residual learning is added to the back-projection layer.Each up-sampling and down-sampling uses the features extracted by all previous up-down sampling corresponding to the value.In the back projection module,global residual learning is added,and each back projection layer will also use the features extracted by all the previous back projection layers.In this way,all features can be connected in a cascade manner,and all feature information generated during the network iteration process can be fully utilized to realize feature reuse and reduce network redundancy,thereby achieving the purpose of improving the image reconstruction effect.Aiming at the mainstream algorithm only extracting image features on a single scale,ignoring the problem of different image feature information at different scales,this paper proposes a multi-scale super-resolution reconstruction algorithm based on dense residual back projection network frame.It uses three different scale channel networks to extract the feature information of the image.According to the different locations of multi-scale information aggregation,the multi-scale model is divided into two types: front-end multi-scale aggregation model and back-end multi-scale aggregation model.Among them,the front end The multi-scale aggregation model is suitable for shallow networks,while the back-end multi-scale aggregation model is more suitable for networks with deep layers.Under different conditions,a suitable aggregation model can be selectively selected for super-resolution reconstruction.The experimental results show that the multi-scale super-resolution reconstruction algorithm based on dense residual back projection network can achieve good reconstruction results.The PSNR values on the Set5,Set14 and Urban100 datasets are higher than the existing mainstream algorithms.