| Image fusion is the key technology to improve the detection level of visible and infrared image sensors.The visible light sensor can receive the reflected light of the object,and the obtained image has clear scene details,high spatial resolution and contrast,but the imaging process is easily affected by severe weather,smoke,aerosol or occlusion,and it is difficult to achieve all-time and all-weather jobs.The infrared sensor performs imaging by receiving thermal radiation in the environment,and the obtained image object is more prominent,but the image texture details are blurred and the contrast is lower.Therefore,the acquisition methods and image characteristics of the visible and infrared image are complementary,and the useful information of the two types of images can be extracted and fused to obtain a fusion image with comprehensive information and good visual effect,thereby further improving the ability to locate,identify and track of the detection system.The sparse representation image fusion method starts from the perspective of sparse representation of image signal,constructs the learning dictionary through the local structure of the image,and uses the dictionary atom to represent the significant features of the image,showing a good fusion performance.However,the traditional sparse representation fusion method still has shortcomings.Firstly,the traditional block-based sparse representation method divides the image into blocks.In the calculation process,the blocks are independent of each other,and the similarity between the blocks is not considered,resulting in the insufficient ability of the dictionary atom to represent the salient feature of images.Secondly,dictionary learning is a large-scale,highly non-convex problem with high computational complexity.In this paper,aiming at the shortcomings of the traditional sparse representation fusion method,the visible and infrared images are taken as the research object,and the non-local similarity of images is used to explore the construction method of image similarity groups,so as to research the sparse fusion method based groups.The main research contents are as follows:(1)By analyzing the the traditional sparse representation fusion method,the advantages and existing problems are pointed out.Combined with the imaging characteristics of visible light and infrared images,the non-local similarity structures of the two types of images are analyzed.Based on the non-local similarity of image blocks,the image similarity group is constructed,and the relation between the local and non-local structure of image blocks is established.In the research process,two methods are used to construct the deformation of similar groups,which provides a new theoretical basis for sparse representation fusion method.(2)Aiming at the problem that the traditional sparse representation does not consider the similarity between image blocks,the visible and infrared images fusion method based similar groups and group K-SVD is proposed.The nonlocal similarity of images is measured by Euclidean distance,and the similarity group is constructed as the similarity group matrix by carring out matrixing process.After that,the improved K-SVD was used to learn the dictionary,and the group sparse decomposition was carried out on the dictionary taking similar groups as units.The fusion rule of the sparse coefficient is to adopt the max-L1.Experimental results show that compared with other fusion methods,the fusion image details are clear and the fusion effect is good.(3)To solve the problem of ignoring the similarity of image blocks and the high complexity of training dictionaries in the process of sparse representation fusion,the visible and infrared images fusion based structured group and double sparsity is proposed.In this method,image structure groups are constructed by vectorization of similar groups.Combining the advantages of analytical dictionary and learning dictionary,the double sparse model of structure group,dictionary training method of structure group and double sparse,and the sparse decomposition of structure group are researched.The fusion rule of the sparse coefficient is to group first and then adopt the Max-L1.Experimental results show that compared with other fusion methods,this method has better fusion effect in both subjective and objective evaluation. |
