Multi-source Image Fusion Technique Via Edge-preserving Filtering

In recent years,the scene multimodal imaging technology has been developed rapidly.However,different types of detectors have different imaging mechanisms.Thus,the information represented by the image is also different in some ways,which results in the information of the scene cannot be descripted through a single image.Therefore,it is an important technology to extract complementary information of multi-source images,remove redundant information and synthesize a composite image which can express scene accurately and completely.Image fusion is an effective solution to this kind of problem.As a branch of multi-source information fusion theory,multi-source image fusion technology has been widely studied,and has been widely used in video monitoring,remote sensing mapping,medical diagnosis,digital photography,and battlefield situation assessment.Under the above research background,aiming at the local block effect,pixel mutation,information loss,spectral distortion,artifacts,and other problems,four types of multi-source image fusion methods including multi-focus image fusion based on Gaussian curvature filtering,infrared and visible image fusion based on multi-level curvature decomposition,remote sensing image fusion based on co-occurrence filtering,and multi-modal medical image fusion based on multi-level edge-preserving filtering are proposed in this dissertation.The main research content of this dissertation can be summarized as follow.(1)Aiming at the phenomenon that the fused image obtained by the traditional spatial domain method is prone to local blocking effect and pixel mutation,a multi-focus image fusion method based on Gaussian curvature filtering is proposed.Firstly,a Gaussian curvature filtering is employed to obtain the feature image through a difference operation between the source images and the filtered images.Secondly,a synthetic focus degree criterion,including spatial frequency and local variance is applied in the feature images to obtain the coarse fusion map.Thirdly,morphological filters and median filters are employed in the coarse fusion map to obtain the optimized fusion map.Finally,the fused image is obtained via a weighted fusion operation.Experimental results demonstrate that the proposed fusion method can effectively overcome the local blocking effect and pixel mutation phenomenon in multi-focus image fusion.(2)Aiming at the phenomenon that the fused image obtained by traditional multi-scale decomposition-based methods is prone to noise and information loss,an infrared and visible image fusion method based on multi-level Gaussian curvature filtering-based decomposition is proposed.This method uses the edge-preserving characteristic of Gaussian curvature filtering and the smoothing characteristic of Gaussian filtering to construct a hybrid multiscale image decomposition model.Firstly,the decomposition model is used to decompose the source images into three different layers: small-scale layers,large-scale layers,and base layer.Secondly,an energy attribute fusion strategy is adopted to merge the base layer,an integrated fusion strategy is adopted to merge the large-scale layers,and a max-value fusion strategy is adopted to merge the small-scale layers.Finally,the fused image is reconstructed through the sum operation of the three fused layers.Experimental results demonstrate that the proposed fusion method can effectively reduce the probability of noise generation and overcome the shortcomings of missing information in the fused image.(3)Aiming at the possible spectral distortion and image distortion in remote sensing image fusion,a multispectral image and panchromatic image fusion method based on cooccurrence filtering and gradient domain pulse-coupled neural network is proposed.In this method,the panchromatic image is decomposed into three layers(coarse-scale layer,finescale layer,and base layer)through co-occurrence filtering.Then,a gradient domain pulsecoupled neural network fusion strategy is employed to merge the base layer and the intensity image of the multispectral image.Finally,the fused image is reconstructed by adding the fused layer,the coarse-scale layer,and the fine-scale layer.Experimental results demonstrate that this method can effectively overcome the spectral distortion and image distortion that may occur after the fusion of multispectral image and panchromatic image.(4)Aiming at the possible artifacts and color distortion in multi-modal medical image fusion,an effective multi-modal medical image fusion method based on multi-level edge-preserving filtering is proposed.This method is combined by hybrid multi-scale transform and neural network-based fusion method.Firstly,the source image is decomposed into fine structure layer,coarse structure layer,and base layer through a multi-level weighted mean curvature filtering and Gaussian filtering.Secondly,the base layer is merged through an energy attribute fusion strategy,Meanwhile,the fine structure layer and the coarse structure layer is merged through a gradient domain pulse-coupled neural network fusion strategy.Finally,the fused image is reconstructed by adding the three fused layers.Experimental results demonstrated that this method can effectively overcome the artifacts and color distortion of the fused image.