Research On Multi-Exposure Image Fusion Algorithm Based On Multi-Scale Decomposition

With the continuous progress of science,technology and living standards,the appreciation of the beauty of human beings is also improved,so the requirements of image quality have become higher and higher.The image quality captured by ordinary imaging devices is poor,a single exposure image can only present the part of scene information,and the sense of visual experience is not strong,which can no longer meet the needs of the public.In order to solve such problems,we can achieve information complementarity by fusing multiple images of the same scene with different exposure levels,thereby generating high-quality images that meet human visual perception.Multi-exposure image fusion(MEF)technology has become one of the main research contents at present,which has widely used in military,meteorological,computer vision and other fields.The existing MEF methods can be divided into five categories: pixel-based,patch-based,multi-scale transform-based,sparse representation-based,and deep learning-based.These methods achieve image fusion by manually designing corresponding fusion rules or feature extraction models.Among them,the fusion methods based on multi-scale decomposition can well preserve the global contrast of the image.By improving the fusion rules,the fusion performance of the algorithm can be effectively improved,the quality of the fused image can be guaranteed,and the running time of the algorithm can be shortened.Based on the methods of multi-scale decomposition,this thesis makes an in-depth study of some problems in the process of image fusion,and puts forward corresponding improved methods.The main innovations are as follows:1)To solve the problem of detail information loss in fusion results,a novel multi-scale decomposition MEF method based on multi-visual feature measurement is proposed.Because the three components(intensity,hue,saturation)in the IHS color space are independent of each other,it is helpful to measure the visual features.Hence,the fusion algorithm is implemented by converting the input exposure images from the RGB color space to the IHS color space.First,a multi-visual feature measurement method is designed to obtained the contrast,saturation and exposure visual features of an image,and an adaptive weighting coefficient is designed for each visual feature.The initial weight maps are constructed by combining three visual features through linear multiplication.Then,according to the value rules,a decision map construction method is proposed to enhance the representation of detail information and optimize the initial weight maps to obtain the middleweight maps.Next,to improve the visual effect of the fused image,the guided filtering is applied to eliminate the noise of the middleweight maps,and obtain the final weight maps.Finally,the fused image of high-quality is generated by image pyramid decomposition and reconstruction.Numerical experimental results indicate that the proposed fusion method is superior to compared methods in terms of subjective visual effect and objective quality evaluations.2)To address the two issues of how to construct the weight maps and extract fine details,a novel multi-scale decomposition MEF method based on edge information measurement is proposed.Since the luminance components can clearly present the structural information of the image,this thesis first obtains the luminance component of the input image by weighted summation.Subsequently,a multi-scale edge information measurement method based on improved multi-scale mathematical morphological filtering is proposed,which is used to extract the edge information of multiple scales from the luminance component,and the extracted multi-scale information are merged by adaptive weighted sum to obtain the edge maps.Next,according to the value rule,the weight maps are constructed by comparing the pixel values of the obtained edge maps.Then,to optimize the weight maps,an adaptive pyramid optimization method based on Gaussian pyramid construction and adaptive guided filtering is proposed,which can obtain edge-preserving smoothing pyramid.Finally,the edge-preserving smooth pyramid and the Laplacian pyramid of input images is combined by multi-resolution hybrid manner,and the fused image is generated by pyramid reconstruction.Numerical experiments conducted on the public datasets show that compared with many compared fusion methods,the proposed method can obtain the better fusion performance.