Reserch On Single Image Super Resolution Based On Sparse Dictionary

It is well known that the sparse dictionary-based image super-resolution(SR)method can effectively input from a single low-resolution image(LR)and reconstruct a high-resolution(HR)image.However,it is found through literature that the general sparse dictionary-based SR method always uses the sparse representation coefficients calculated in the low-resolution dictionary as the high-resolution sparse representation coefficients directly,Then,it uses these coefficients and the high-resolution dictionary to obtain high-resolution images.However,it is found through experiments that the high and low resolution sparse representation coefficients are different,so it is inappropriate to directly use the sparse representation coefficient calculated in the low resolution dictionary as the high resolution sparse representation coefficient.At the same time,by observing the reconstructed images obtained by some SR methods,it is found that although many reconstructed images have obtained high-resolution images,the edges of the images are more or less blurred.Therefore,this article studies the above two issues.The main research work is as follows:(1)A sparse coefficient optimization method based on neural network is proposed to train the sparse representation coefficient matrix of the low-resolution dictionary in the dictionary reconstruction process and the sparse representation coefficient matrix of the actual high-resolution image to predict its error.Compensating the error makes the sparse matrix in the given image super-resolution reconstruction process closer to the actual value,in order to get a better reconstructed image.(2)A new and simple 'catch doll' classification method is proposed.This method simulates the process of catching dolls in shopping malls to quickly and easily classify the data,instead of performing poorly during the sparse representation coefficient classification neural network training K-means classification method to reduce the time required for network training and improve the reconstruction effect.(3)A single image super-resolution method based on the structure tensor constraint and the down-sampling constraint of the reconstructed image is proposed.The previous image reconstruction constraint term based on the structure tensor is improved.Sampling the constraint term,a comprehensive edge constraint optimization term is proposed,and the reconstructedimage is constrained and optimized for reconstruction again,so that the pixels near the edge of the reconstructed image after constraint optimization reconstruction are clearer,and the pixels on both sides of the edge are effectively distinguished.To reduce blurring at the edges of the image.(4)Utilizing some high-resolution test images and SET5 and SET14 high-resolution test image sets,a neural network-based sparse coefficient optimization method and a single image super-resolution method based on structural tensor constraints and reconstructed image down-sampling constraints are proposed.undergone an experiment.The results show that the proposed two new SR methods achieve good image reconstruction results and higher peak signal-to-noise ratio(PSNR).The proposed classification method can effectively reduce the training time of the neural networks after the conventional K-means classification method,and obtains good results.