Research On Image Super-resolution Reconstruction Based On Convolutional Sparse Representation

In real life,due to the limitations of imaging equipment physical characteristics,people usually cannot get the ideal clear image.However,the cost to improve the resolution of the image by improving the hardware configuration of the imaging system is very high,and in a short time it is difficult to overcome certain core technical problems.Therefore,it is of great value to improve the resolution of the image by using software and algorithms.Image super-resolution reconstruction technology uses digital image processing methods,from a single or multiple low-resolution images to reconstruct a corresponding high-resolution image.This paper focuses on how to use the prior information of a low-resolution image in the current scene to predict the highfrequency information lost by the image,and reconstruct the corresponding clear image.At present,most of traditional image reconstruction methods based on reconstruction and learning are based on the image block.However,these methods ignore the correlation of the pixels in the image,so that the reconstructed image has obvious block effect,which leads to the degradation of the reconstructed image quality.The image restoration method based on the convolutional sparse coding model is not affected by the image size,and directly processes the image,which overcomes the defects of traditional model based on the image block.At the same time,the convolution filter with translational invariance,which can better represent the image features.This paper,based on convolutional sparse coding,proposes two different image superresolution reconstruction methods by combining super-resolution reconstruction methods that based on reconstruction and learning.The main contents and contributions of this paper are as follows:First,image super-resolution reconstruction method based on leaning-based convolutional sparse coding.The model utilizes the convolutional sparse coding technique to extract features and reconstruct the whole image,avoiding the effects of block effect in image reconstruction.At the same time,this model can come true image super-resolution reconstruction in the case of a low-resolution image,greatly reducing the time complexity of the model training.In the training process,considering the main difference of the high and low resolution images lies in the high-frequency information of the image,makes the different information between reconstructed image and the corresponding degraded image(image using the same degradation model)as the highresolution training image,and takes high-frequency information of corresponding degraded image as the low-resolution training images.Then use the iterative shrinkage threshold algorithm to get high and low resolution filters and convolutional sparse representation coefficients,finally the mapping relation between the high and low resolution sparse coefficients is obtained by the least square method.In the process of reconstruction,use the low-resolution filters and the mapping relation between convolutional sparse coefficients to obtain high-resolution image sparse representation coefficients,then use high-resolution filters and sparse coefficients to reconstruct the difference of original high and low-resolution image,so as to reconstruct the original high-resolution image.This method makes full use of the image information,and provides more effective information for image reconstruction,thus improving the quality of reconstructed image.Second,image super-resolution reconstruction method based on non-local selfsimilar convolutional sparse coding.This model uses image nonlocal self-similarity and convolutional sparse coding technology to come true image super-resolution reconstruction.Regard the reconstructed image as a noisy image,and the corresponding sparse representation coefficients also have noise,by making the image reconstruction problem into a minimized convolutional sparse coding noise to obtain convolutional sparse representation coefficients(feature map)that are closer to the sparse representation coefficient of the original clear image.In the process of reconstruction,the non-local average denoising method is used to modify the current high-resolution image,and then the convolutional sparse coding technique is used to obtain the convolutional sparse representation coefficients of the modified image.Finally use an iterative threshold algorithm to minimize the convolutional sparse coding noise,obtaining the high-resolution convolutional sparse representation coefficients.Continuous updating the iteration,the final high-resolution convolutional sparse representation coefficients can be received,and the high-resolution image can be reconstructed by using the convolutional sparse coding theory.