Infrared And Visible Image Fusion Based On Saliency Analysis And Hierarchical Joint Low-Rank Representation

Infrared and visible image fusion,as an important branch of image fusion research,is significant in reality and has been applied to widely in many fields.Its main goal is to not only completely extract the thermal target information of the infrared image,but also retain the detail information of the visible image.At present,sparse representation and low-rank representation theory have been successfully applied to image fusion,and some fusion results have been achieved.However,these image fusion methods usually only encode each image patch in isolation,and then determine the importance of the image patch according to its representation coefficient,ignoring the correlation between image patches and without considering the global target characteristics of local image patches.So the fusion results can not accurately highlight the global "interested" targets of the source images and lost some other image information.However,image saliency analysis technology,to a certain extent,can simulate human visual attention mechanism and estimate the target regions in the images.Based on this,a new infrared and visible image fusion algorithm is proposed based on saliency analysis and hierarchical lowrank representation.(1)Image saliency analysis is combined with image fusion,and a hierarchical joint low-rank representation model is presented in this paper.The feature similarity between image regions is considered in this model.And the model decomposes the input images and can effectively extract the globally significant information in the source images,which facilitates the subsequent fusion process to integrate these salient object information.(2)A background dictionary is constructed for the hierarchical joint low-rank representation model.Different from the traditional low-rank representation model using original data as the dictionary,the background dictionary constructed in this paper has better reconstruction abilities for background region signals,and the dimension of the dictionary is lower,which can effectively reduce the computational complexity of the fusion algorithm.(3)Aiming at the problems of existing image fusion algorithms based on representation theory mentioned above,a novel infrared and visible image fusion algorithm is proposed based on saliency analysis and hierarchical low-rank representation.The algorithm mainly includes the following steps: a)dividing the input images into several fixed-size nonoverlapping image patches;b)constructing hierarchical tree structure and background dictionary for the model;c)using the hierarchical joint low-rank representation model proposed in this paper,to encode the image patches and obtain low-rank representation coefficients;d)designing the fusion rules to fuse each part and finally reconstructing the fused image.(4)Compared with the existing nine representative fusion algorithms,the fusion results are comprehensively analyzed in terms of subjective evaluation and objective evaluation indexes.The experimental results demonstrate that the proposed method can obtain better fusion results,especially for preserving and highlighting the salient target information of the source images.And the fusion results have higher contrast and better visual fidelity.