Learning Sparse Representation Model For Thermal And Visible Image Fusion

As a fundamental problem in computer vision community, the thermal and visible image fusion uses some algorithms to extract image features from different wavelengths, and fuses them together. The result fused image, as compared with originate images,has more clear and more comprehensive information, therefore, more convenient for human to identify and to understand the scene. Thermal and visible light work in different bands, so they have different imaging principles and characteristics.By fusing the thermal and visible images,complementary information contained in both image modals can be put together into one image and possible drawbacks result from using single image modal can be eliminated. Nowadays, the thermal and visible image fusion has been applied in many areas including video monitoring, military and power system. With the development of related technologies, the application field of multi-modal fusion is becoming more and more diversified. However, more challenges still exist in the multi-modal image fusion.At present, many algorithms have been proposed to fuse different modal images. However, there are more or less some problems in these proposed algorithms. In order to improve the quality of multi-modal image fusion, in this paper we present an algorithm based on Laplace sparse representation model to fuse images of different modes. The main contributions of our work are:(1)Because of the different imaging principle of different modal images, the traditional registration algorithm for single modal images cannot reach an ideal effect in different modal image registration. Therefore, the correspondence between different modals is difficult to accurately established. To solve this problem, a multi-modal image registration algorithm based on distance metric learning is proposed in this paper. Learning the distance metric between different modal images can accurately establish the transform relationship between the different modal images.(2)Based on the existing multi-modal image fusion algorithm, the sparse representation combining with a Laplace regularization is introduced to improve the robustness and the quality of image fusion result. It consists of three parts:dictionary learning, sparse representation and multi-modal fusion. Experiments show that the proposed algorithm can achieve a better image registration and the fusion results are improved compared with other algorithms.