Research On Super-resolution Algorithms Based On Self-similarity And Sparse Representation

With the rapid development of science and technology in recent years,especially the promotion of Internet communication and storage technology,it is possible to transfer image files that containing a large amount of data.With the People's increasing demand for image quality and an important evaluation index of image quality,the resolution of image has been paid much attention by researchers.How to improve the resolution of image and improve the quality of image has become one of the hot research topics in the field of image processing.At present,scholars divide the super-resolution algorithm into three categories:interpolation-based,reconstruction-based and learning-based.Due to the introduction of a lot of prior knowledge,the learning-based algorithm is obviously superior to other algorithms in image reconstruction,so it has become a hot research topic in super-resolution algorithm.In order to improve the effect of super-resolution algorithm in a further step,the super-resolution algorithm based on self-similarity and the super-resolution algorithm based on sparse representation are studied in this paper.In order to solve the problems of self-similarity algorithm,a partial improvement have been made,and then a super-resolution algorithm based on self-similarity and sparse representation is studied by combining the improved self-similarity algorithm with sparse representation algorithm.The main methods and conclusions are as follows:(1)The super-resolution algorithm based on self-similarity is studied.Because the traditional gaussian pyramid uses a single sampling method,fewer layers can be created,so the original image block can find fewer similar blocks in gaussian pyramid,which seriously affect the quality of image reconstruction.Based on this,an improved gaussian pyramid is adopted in this paper.In addition to the traditional transverse and longitudinal simultaneous sampling,the improved method increases the sampling only in the transverse direction and the longitudinal direction,and the improved gaussian pyramid can obtain more similar blocks when searching for similar blocks.In the process of similar block search,the traditional algorithm adopts fixed threshold,which is not ideal for different images.Based on this,an adaptive threshold is studied in this paper.the algorithm constructs different adaptive threshold for different input images,which improves the reconstruction effect.The improved self-similarity algorithm uses multi-dimensional Gaussian pyramid to construct sub-sampling layer,and then uses adaptive threshold to find similar blocks.Finally,the weight is constructed according to the distance of similar blocks.The experimental results show that the improved algorithm has a better effect than the traditional self-similarity algorithm.(2)The theory and method of sparse representation are studied,and the shortcomings of self-similarity algorithm and sparse representation algorithm are analyzed in detail.The self-similarity algorithm only uses the global information of the image,but does not introduce the local information and the external information.The reconstructed effect of the traditional sparse representation algorithm depends on the selection of training data and the dictionary needs to be trained ahead of time.To overcome the disadvantages of the two algorithms,a super-resolution algorithm based on self-similarity and sparse representation is studied in this paper.First,the algorithm uses the self-similarity of the image to construct the high-resolution and low-resolution image block samples,and then trains the samples by the way of joint training to obtain the high-resolution dictionaries and-low-resolution dictionaries.Finally,the sparse vector is solved based on the input low-resolution image and the low-resolution dictionary,and then the sparse vector is combined with the high-resolution dictionary to get the final high-resolution image.The effect of this algorithm is independent of training data,which solves the disadvantage that traditional algorithm depends on training data.Furthermore,this algorithm does not need additional training dictionary,and realizes the online updating of dictionary.The experimental results show that the proposed algorithm has good performance both in subjective and objective evaluation.