Research On Image Fusion Based On Joint Feature Extraction

With the expansion of the application of image fusion imaging,fusion objects have developed in a diversified direction.The information differences between multi-sensor images and the information redundancy between identical sensor images bring a series of difficulties and challenges to image fusion.How to establish an effective and robust method of information extraction and fusion to the fusion task is a subject that needs to be studied urgently.This dissertation focuses on the problems of multi-sensor and identical sensor image fusion.In this dissertation,sparse representation and subspace representation are used to solve the above problems.The research contents are as follows:1.The problem of spatial inconsistency in extracting features of multi-sensor images in traditional sparse representation is investigated,a joint convolutional sparse representation model is proposed.The model combines convolutional sparse representation and joint sparse coding to realize joint convolutional sparse representation.It can extract common and unique features between multi-sensor images in a global sparse form,where the global sparse form ensures the spatial consistency of feature extraction.Experiments show that compared with traditional sparse representation,the model has good performance in the global feature extraction,thereby ensuring the spatial consistency of feature extraction.2.The problem of details loss in extracting features in identical sensor images in traditional subspace representation is considered,a joint low-rank representation model is proposed.This model constructs a joint low-rank representation through a low-rank representation,and it can represent and extract the common principal components between identical sensor images by sharing the low-rank part.At the same time,the model uses the sparse part to jointly represent and extract the unique features of the identical sensor images.Experiments show that,compared with traditional subspace representation,the joint low-rank representation has good performance in feature extraction accuracy,thereby reducing the loss of feature details.3.In order to deal with the infrared intensity decrease of local targets and scene details blur in the fusion of infrared and visible images,an image fusion method based on joint convolutional sparse representation is proposed.First,the method uses the joint convolutional sparse representation to extract the unique features of infrared images(target intensity)and the unique features of visible images(scene details)in a global sparse form.Secondly,the extracted features are reasonably fused by designing parameterized weight fusion rules to achieve the final image fusion.Experimental results show that this method has better performance than traditional sparse representation fusion methods in fusing the local target intensity of the infrared image and preserving the details of the visible scene.4.In order to deal with the problem of blurring at the boundary of the focus area in multi-focus image fusion,a multi-focus image fusion method based on joint lowrank representation is proposed.First,the method introduces the common low-rank part and the sparse part in a joint low-rank form to extract features of the defocused and focused region between the multi-focus images.Secondly,by designing local measurement rules to obtain the fusion decision map and further to guide the final image fusion.Experimental results show that this method is superior to traditional multi-focus image fusion methods in solving the blur problem.