Research On Super Resolution Of Single Image Based On Sparse Representation

In recent years,among the single-image super-resolution reconstruction algorithms,the sparse representation-based image super-resolution algorithm has attracted the attention of scholars and researchers due to its good visual effect in reconstructed images.However,this method also has certain problems in the process of image super-resolution reconstruction: First,this type of algorithm does not make full use of the texture structure information of the image itself,so that there are artifacts and block effects in the reconstructed image,and the image reconstruction process is time-consuming.Second,this type of algorithm only uses external databases for dictionary learning,ignoring the internal information contained in the image itself,resulting in poor visual effects at the edges of the reconstructed image.Aiming at the above two problems,this paper proposes the corresponding image super-resolution reconstruction algorithm.The algorithms and innovations proposed in this paper are as follows:1.In the super-resolution reconstruction process of sparse representation,the input image needs to be divided into overlapping blocks from top to bottom.However,this process does not consider the structural characteristics of the image itself,and directly divides the smooth part and the edge part of the image into image blocks of the same size,which makes the cost of time is high,and there are artifacts at the edge of the reconstructed image.To solve this problem,in the third chapter,this paper proposes an image super-resolution algorithm based on block adaptation.The main contribution of the algorithm is to improve the image block method.Based on the image quadtree block idea,according to the texture structure information of the image itself,the adaptive block algorithm is used to block the image and divide the smooth part of the image into large-scale image blocks,the edge of the image is divided into small-scale image blocks.For the two types of image blocks after division,we choose two different methods for reconstruction.The smooth image blocks contain less high-frequency information,and the reconstruction is directly performed by bicubic interpolation,which can avoid the seam effect and the block effect,and the time consumed for reconstruction is almost zero.For the edge image blocks with prominent texture structure,the sparse representation on the over-complete dictionary is used for reconstruction.Then two types of reconstructed image blocks are placed at the corresponding positions of the target high-resolution image to complete the super-resolution reconstruction of the image.Finally,the simulation is carried out on the MATLAB experimental platform.Experimental results show that compared with Yang's algorithm and NESR's algorithm,the proposed algorithm has the best visual effect on reconstructed images,and its reconstruction time is about half of NESR's algorithm.2.In the reconstruction algorithm based on learning dictionary,the quality of the reconstructed image largely depends on the correlation between the input image and the training samples.However,in the super-resolution algorithm based on sparse representation,only external data is used to construct training samples,and then dictionary learning is performed.This process does not consider the mining of the input image's own information.Therefore,for some input images with weak similarity to the training sample images,due to the limited ability of dictionary expression obtained by external data training,it is difficult to accurately provide the high-frequency information required for the reconstruction process,as a result,high-quality images cannot be reconstructed.To solve this problem,in the fourth chapter,a classification-based image super-resolution reconstruction algorithm is proposed.The core idea of the algorithm is to use the image similarity judgment method to divide the input images into two categories: structural similar images and structural dissimilar images.For structural similar input images,the image contains a large number of self-similar image blocks,so we use a super-resolution algorithm based on self-learning for reconstruction,and the reconstructed image has better visual effects.For structural dissimilar input images,that is,the texture structure of the images is relatively complex,reconstruction requires more high-frequency information.Therefore,we propose an image super-resolution algorithm that combines self-learning and sparse representation,and simultaneously reconstructs it using external databases and the internal information of the image itself,which can better recover its lost high-frequency information.The algorithm proposed in this chapter selects a targeted reconstruction method for different types of input images,so that it can achieve a good reconstruction effect for different types of input images.For images with complex texture structures,the reconstruction quality has been significantly improved.Comparing the algorithm of this chapter with the algorithm based on sparse representation and the algorithm based on self-learning,the experiment shows that the PSNR value of the reconstructed image of the algorithm proposed in this chapter is improved by 0.4 ? 0.7d B under the condition of 3 times super resolution.