Research On Multi-focus Image Fusion Based On Sparse Representation With Over-complete Dictionary

Generally a single image cannot reflect all the information of one scene accurately and comprehensively. The technology of image fusion can improve the accuracy and incredibility of the image description through combining the complementary information and removing the redundant information of the pictures from the same scene, so that a new image about the scene is obtained which is suitable for understanding and analyzing. As a branch of image fusion, the purpose of multifocus image fusion is to make the targets of the scene clear and homogeneous.In this thesis, the theory of sparse representation with over-complete dictionary are studied firstly. Then the theory is introduced into fusion technology field and new image and information fusion schemes are presented.The main contributions of the research are summarized as follows:1. Firstly, the sparse representation theory and its applications in the domain of image processing and analysis have been elaborated. Secondly, the steps of K-SVD dictionary training method have been listed and three over-complete dictionaries have been accomplished via this method. Finally, the steps of OMP algorithm have been given and meanwhile we have studied the way to reset the residual error.2. Above the image fusion algorithm using sparse representation previously presented, we put forward a new method in which a new fusion rule is adopted. In this rule the sparsity of the coefficient vector has been considered as the active factor. According to the active factor then we can choose the coefficients to construct the fusion image. The experiments show that the new rule is effective and superior to the same class of previous rule presented in several indexes.3. Totally considering the characteristic of the multifocus image fusion we present a new scheme based on blocked sparse representation using distinct window. To remove the block effects, we choose a weighted average rule in which the weighted factors have been calculated with the l1norm to fuse the coefficients. Meanwhile we have discussed the effects the over-complete dictionary has taken to the quality of the final fusion image.