Adaptive Image Fusion Based On Sparse Representation

Image fusion is to synthesize two or more images to an image, which makes thefused image more suitable for human visual characteristics and the further analysisand understanding of computer, the images are taken from different sensors at thesame or different time for a particular scene on the image or image sequenceinformation.Based on the state-of-the-art image fusion methods, we proposed a new sparserepresentation-based image adaptive fusion algorithm. First, the two source imagesare represented with sparse coefficients using the overcomplete dictionary whichtrained by the K-SVD algorithm. Then the coefficients are combined with theadaptive fusion policy which depends on the characteristic of the coefficients. Finally,the fused image is reconstructed from the fused coefficients and the overcompletedictionary. The algorithm can denoise effectively during the process of sparserepresentation, and also could overcome the shift variance problem. The fused imagehas a lot of advantages such as high resolution, the rich image information and highrobustness, etc. It is widely used in the areas such as pattern recognition, medicaldiagnosis, intelligent navigation and so on. The main research contents aresummarized as follows:(1) The process of image registration during the preprocessing of the imagefusion was introduced. Several transformation models of image registration and somemethods on general image registration were mentioned in detail, and also given thedetailed process of image registration based on the Harris corner detection, finallyproved the effectively of the algorithm by the experimental results.(2) Through the in-depth study of the theory on the sparse representation, weintroduced detailedly of the method of the dictionary construction based on the K-singular value decomposition (K-SVD), and then we proposed a new algorithm ofover complete dictionary training with multiple source images, and given thecorresponding experiments and the analysis of the results.(3) Common image fusion methods based on spatial domain and transformdomain were discussed. After the analyzing and comparing the results of the fusion,we proposed a new method of adaptive image fusion based on sparse representation.That is, the fusion strategy which connected with the variance of sparse coefficientand the sparse degree. Furthermore, the effects of image block size and the globalerror during the process of the sparse coding were studied. And the sliding windowtechnique was adopted to solve the translation invariance in the process of fusion.Finally, we made a research on the denoising performance of the image fusion methodbased on sparse representation, and made comparisons with other well-knownalgorithms.