Image Fusion Algorithms Based On Sparse Representation

In order to eliminate the redundancy and exploit the complementary of different images from a same scene, the image fusion technology is proposed. Image fusion technology integrates multiple images to a single image, which can illuminate objects better. The main progresses of image fusion are composed of feature extraction, feature fusion and reconstruction. As the signal can be represented sparsely by a small number of atoms, the theory of sparse representation has been widely used in the field of signal processing. In this paper, image fusion methods based on sparse representation are studied.As the methods of learning dictionary can adaptively and sparsely represented the images, the MOD and K-SVD methods of learning dictionary are deeply analyzed and applied to image fusion. In addition, with the development of deep learning, the Sparse Auto-Encoder(SAE) has been proposed and widely applied in the field of image processing, and its performance of sparse representation is also satisfied. Thus, a novel image fusion method based on SAE is proposed. Firstly, the dictionary is obtained by SAE; the acquired dictionary is then used to represent the image sparsely and obtain the sparse coefficients; finally, the sparse coefficients are fused and reconstructed to the fused image. In order to analyze the performance of MOD, K-SVD and SAE methods of learning dictionary, some experiments are conducted.Furthermore, as the natural image is composed with different components, an image fusion method based on cartoon and texture sparse decomposition is studied. Firstly, the cartoon components and texture components of input images are extracted using Morphological Component Analysis with Total Variation(MCA-TV) algorithm. In this step, the cartoon components and texture components can be sparsely represented by Curvelet and local DCT transformation, respectively. The sparse coefficients corresponding to cartoon components and texture components are then obtained. After that, these sparse coefficients are fused using different fusion rules, respectively. At the next step, the fused cartoon component and texture component are acquired by transforming the corresponding fused coefficients inversely. Finally, the superposition of the fused cartoon and texture components is modeled as the fused image. To illuminate the efficiency of the studied fusion method, several experiments are executed. The experimental results show that the studied method achieves better quality of fused image than the existing state-of-the-art methods.