Research And Application Of Infrared And Visible Image Fusion Algorithm Based On Low-Rank And Sparse Representation

Information fusion technology is developed to improve the comprehensiveness and accuracy of image application.It can obtain more useful and complete information content by integrating multi-source images in the same scene.Infrared and visible image fusion technology is rapidly developing,which can combine the excellent resolution of visible and infrared images with rich texture information.This technique can effectively enhance the effect of image application.The algorithm fusion framework based on the low-rank sparse representation method can effectively retain the global information and detailed texture information of the source image,which is convenient for later application.In terms of lowrank sparse representation,this paper explores effective fusion methods of low-rank layer and sparse layer,so as to achieve high-quality subjective and objective effects of the final fusion image.The main research of this subject is as follows:(1)A new hybrid fusion model is proposed,which organically combines RPCA(robust principal component analysis)and RTV(relative total variation)technologies.The source image is decomposed into low rank layer and sparse layer by using RPCA technology.For the low-rank layer containing a lot of structural information and global information,the method based on relative total variation is used for fusion.The relative total variation decomposition can effectively separate the structure and texture.For the fusion of sparse layers with background information and noise,the average energy method is used to preserve the useful background and remove the useless noise.Compared with the other five fusion methods,this method has better effect.From the perspective of the seven evaluation indexes of image fusion,the fusion effect of this method also has certain advantages.The average gradient,mutual information and edge intensity are greatly improved,and the fusion effect of multi-target and multi-detail source images is more obvious.(2)Considering that the image has shape and scale parameters,generalized maximum correntropy criterion(GMCC)is introduced into the local similarity measurement of GoDec algorithm to get a robust low-rank sparse decomposition algorithm.After the source image is denoised by the improved GoDec decomposition,the low-rank layer is decomposed into high and low frequency sub-bands by the non-subsampled contourlet transform(NSCT).The low frequency is fused by the Bayesian hierarchical model to maintain the overall contour of the source image,the high frequency is fused by the phase consistency criterion to maintain the details,and the sparse part is fused by the weighted average strategy to preserve some details.Compared with the other five algorithms,the average gradient and edge information evaluation factor of the fusion results of this method are improved in the order of 29.5%77.8%,edge intensity is improved in the order of 24.4%-77.2%,spatial frequency is improved in the order of 35.6%-86.7%,and fusion image definition is improved in the order of 36.1%78.4%.(3)Design and develop an image fusion system,which includes three parts:source image selection,image fusion and calculation index.The system selects an appropriate fusion algorithm to fuse the source image and displays it.The objective evaluation index of the calculated result is used to observe its effect.The proposed algorithm is applied to pedestrian target detection task to improve the accuracy of target detection.