Single Image Super Resolution Based On Compressed Images

Super-resolution reconstruction of compressed images refers to obtaining high resolution clear images from low resolution compressed images by certain technique.It can restore the details of the image lost in the compression process and remove the obvious artificial traces in the compressed image.The super-resolution reconstruction method of compressed image proposed in this thesis is a learning based reconstruction method.The main research results are as follows:1.A super-resolution reconstruction method of compressed images based on subspace regression is proposed,which can reconstruct images in the same scale.It mainly includes two parts: training and testing.In the training stage,different subspaces are first trained to classify the image blocks.Then,Extreme Support Vector Regression(ESVR)is applied in different classes to train nonlinear regression models from low resolution image blocks to high resolution and high frequency blocks.In the testing phase,the multiple extreme support vector machine regression fusion method is used to estimate the high frequency.The selection of multiple regression models is based on the representation coefficient of the image block on all subspaces.Finally,the estimated high frequency is taken as a constraint to reconstruct the result graph,and the result graph is smoothed to remove the shock effect further.The experiment shows that the average PSNR increases 0.15 dB for the same test set compared with algorithm A+.2.A super-resolution reconstruction method based on vehicle image is proposed.According to the degradation process of vehicle image,the reconstruction algorithm is divided into three parts:(1)The compression image reconstruction algorithm in this paper is used to remove the compression effect in the vehicle images.(2).The scale up super resolution reconstruction is carried out for the decompressed result graph.In the training stage,different subspaces are first trained to classify the image blocks.Then,ESVR is applied in different classes to train models from low resolution interpolated image blocks to high resolution and high frequency blocks.In the test phase,the strategy of multiple ESVR fusion is adopted to obtain the final result.(3)Blind image deblur is performed for vehicle image with scale expansion.Firstly,blur kernel is deduced,and then the final result is obtained by image deconvolution based on the estimated blur kernel.The experiment shows that the average PSNR increases 0.24 dB compared with algorithm A+.