Research On Multi-exposure Image Fusion Method For HDR Image Synthesis

In the natural scene,it is a common method to fuse multiple images with different exposure levels to obtain high dynamic range(HDR)images.These methods can often obtain satisfactory results in static scenes.However,in real life,there are inevitably moving objects and other factors,the final fusion results may cause ghosting artifacts.In most of the methods,serious geometric distortion and color artifacts will be introduced in the fusion process due to the interference of moving objects when dealing with the foreground moving region.The mechanism of visual attention is essentially ignoring irrelevant information and focusing on important information.In the process of image fusion and deghosting,the moving region is the key information.It can guide the iterative fusion in the process of ghost removal and reduce the interference of the moving region as much as possible to improve the performance of the multi-exposure image fusion(MEF)method.Therefore,this thesis focuses on the related characteristics of attention mechanism and its application in dynamic scenes,and proposes two multi-exposure image fusion methods.Firstly,a new attention-guided neural network(ADeep HDR)is developed based on the codec network.Different from the previous approach,this thesis uses an attention module to guide the process of image merging.The attention module mainly considers the factors affecting image fusion from the perspective of motion region detection.In addition to detecting the edge detail feature of the moving region,it can also extract salient features under muitiple color channels in different exposure sequences.Then the extracted feature vector is used as the input of the neural network to guide the image merging process to remove the ghost.For the merging module,this thesis tries different sub-network variants to make full use of the hierarchical features.In addition,fractional order differential convolution is also used in the sub-network to extract better edge detail features.The method in this thesis is validated and analyzed in Kalantari,Hu,Ma,Sen and other extended datasets.The results show that it can avoid the artifacts caused by wrong optical flow estimation and large foreground motion better.Moreover,this thesis also innovates on the traditional non-deep learning image fusion method,and improves a multi-exposure image fusion method using the independent component analysis(ICA).Different from other traditional methods,the proposed method firstly decomposes the input image into a series of structural patches,and then uses fast independent component analysis in multi-color channels(Color-Fast ICA)to detect the moving regions and salient features of the input exposure sequences to obtain the structural consistency map.Finally,the patch-based optimization strategy is used for iterative fusion of each patch,and the final fusion result is obtained.In this thesis,a large number of qualitative and quantitative experiments are carried out on different test sets.The results show that the proposed algorithm is superior to most advanced multi-exposure image fusion methods and can obtain better and clearer high-dynamic-range images.To sum up,this thesis introduces the two proposed multi-exposure image fusion methods,and starts from the relevant characteristics of visual attention mechanism to elaborate its research and application in deep learning and non-deep learning dynamic scene image fusion.Finally,the multi-exposure image fusion methods for HDR image synthesis are prospected,and the shortcomings of the existing methods and the direction of improvement in the future are also discussed.