Research Of Image Fusion Algorithms Based On Guided Filtering And Total Variational

Image fusion technique can combine two or more images to an image,complement unique features and eliminate redundant information between each other,so as to achieve a more accurate and comprehensive description of the scene.Infrared(IR)and visible(VIS)image fusion and medical image fusion are two important branches of it.IR and VIS image fusion can generate the visible information effectively,meanwhile it can offer nighttime information which is imperceptible to the human eyes.Medical image fusion can combine the multimodality medical image information efficiently which is very helpful for clinical diagnoses and treatment.This paper presents the fusion approach of IR and VIS images and the fusion method of medical images,the main research content includes:1.As to the problem of spatial inconsistency existed in the traditional IR and VIS image fusion methods,a novel two-scales image fusion method based on guided filter is proposed.First,the IR and VIS images are decomposed with a two-scale average filter to generate the base layers and detail layers.Second,phase congruency(PC)is applied to generate the saliency maps of detail layers,and the sum modified laplacian(SML)is adopted to generate the saliency maps of detail layers.Finally,the guided filter is applied to construct the weight maps of base and detail layers,and the resultant image is reconstructed by adding the base and detail layers.2.As to the problem of edge information drop-out existed in the traditional medical image fusion methods,a novel image fusion approach is proposed based on the non-linear structure tensor and variational method.First,a new energy functional is defined in spatial domain.Second,the gradient approximation item is obtained based on non-linear structure tensor,and in order to construct the non-linear weight maps of structure tensor,we utilize the rolling guidance filter to retain the edge features of source images,and applied the SML operator to the filter output image to generate the saliency maps.Next,the fidelity item is acquired by fusion algorithm based on nonsubsampled shearlet transform(NSST).Finally,the gradient descent method is proposed to solve the unconstrained optimization problems and obtain the fusion result.Experimental results show that the proposed methods not only preserve the details of source images but also suppress the artifacts effectively,and the spatial continuity of fusion results can be significantly improved.We have compared our methods with some popular multi-scales fusion methods and guided based methods,our approaches have a better performance in both subjective visual and objective evaluation,which indicates the reasonability and efficiency of our methods.