Research On Image Fusion And Super-resolution Reconstruction Based On Discriminative Dictionary Learning And Convolutional Sparse Representation

Image fusion technology is an important branch in the field of digital image processing,which has attracted more and more researchers' attention,and has made significant research progress.Among them,the image fusion method based on sparse representation has been widely concerned by researchers because of its excellent performance.However,most of these methods use a dictionary to represent the entire image,and cannot effectively express different morphological structures in natural images.And these methods assume that the source image is clear and high-resolution.Obviously,it is difficult for the source image actually obtained to meet this condition.In addition,the traditional sparse representation method divides the image into stacked blocks and pulls the image blocks into vectors for processing.In this process,the consistency between the overlapping blocks is ignored.Aiming at these problems,this paper proposes a joint implementation method of image fusion and super-resolution reconstruction based on discriminative dictionary learning to effectively improve the quality of the fused image.The specific methods are as follows:(1)In order to overcome the limitation of the expression ability of a single dictionary,the complementary representation mechanism of analysis and synthesis of sparse representations is integrated into dictionary learning,and a pair of analysis and synthesis dictionaries with strong expression and discrimination capabilities are trained to express images.Structural and edge detail components,as well as image background and brightness components.In addition,in order to improve the salient feature retention performance of the fused image,a fusion strategy based on the salient features of the image block is proposed for the fusion of the main structure and edge detail components of the fused image.This method can better save image information and improve image contrast.(2)For low-resolution image fusion,a method based on discriminative dictionary learning and advantage embedding to realize image fusion and super-resolution reconstruction is proposed.First,jointly train two pairs of low-rank,sparse dictionaries and a conversion dictionary.One pair of dictionaries is used to represent the low-rank and sparse components of the input image,and the other pair is used to reconstruct the high-resolution fusion of low-rank and sparse components.To establish the potential relationship between high-and low-resolution images.In order to improve the visualization of the fusion and super-resolution reconstruction results,a structure information compensation dictionary was learned to compensate for lost detailed information.In addition,in order to integrate the advantages of the excellent image fusion method into the fusion result,a deconvolution-based advantage embedding scheme is proposed.The actual experimental results show that the proposed method can effectively enhance the visual effect of the fusion image.(3)To solve the problem of sparse representation and vectorization of traditional sparse representation methods to destroy the consistency of adjacent pixels,a joint implementation method of image fusion and super-resolution reconstruction based on convolutional sparse representation is proposed.This method uses a set of sparse feature maps and filters to represent the entire image,so it can effectively maintain the correlation of local features of the source image.A high-and low-resolution filter joint learning framework is designed.The low-resolution filter is used to decompose the input image to obtain a sparse feature map,and the high-resolution filter bank is used for super-resolution reconstruction of the image.And according to the morphological characteristics of different components,different fusion rules are designed.Finally,experiments show that the proposed method can improve the image contrast while effectively retaining the salient features of the image.