Research On Image Fusion Based On Regional Segmentation

Image fusion, especially the heterogeneous sensor image fusion, is a branch of multi-sensor information fusion. In recent years, it has been more and more widely applied in the military target recognition, medical pathology diagnosis, urban planning management etc.As there are many kinds of methods, this paper pays great attention to the region, mainly researches and discusses the fusion methods based on regional segmentation.The basic theory of image fusion has been introduced firstly, including the concept, steps,levels and the characteristics of commonly used methods. Then, we studied the theory of image segmentation, mainly on methods based on edges and regions.The method of image fusion based on regional segmentation is performed on pixel level.The traditional technique uses the average gradient to judge image clarity, but it cannot measure the local clarity exactly, and it also needs more time. The Energy of Laplacian(EOL)method has a better effect on the clarity judgment, and this kind of algorithm is simple, only involves "plus" process, can shorten the operation time. In this paper EOL is used to replace the commonly applied average gradient metric. Experimental results demonstrated that the proposed method was effective. In addition, experiments on infrared and low-light images also achieved good results.The algorithm of regional segmentation image fusion based on similarity characteristics fused the source images from three aspects, luminance, contrast and structure comparisons. As the traditional similarity parameter can not fit the structure information well, this paper induced wavelet similarity, which reflects the edges information accurately, to substitute traditional correlation coefficient similarity. In addition, this paper adopted contrast sensitivity function reflecting the human vision to weight the wavelet coefficient on different scales,divided redundant and complementary regions by weighted similarity. Experimental results demonstrate that the proposed method outperformed the traditional fusion algorithm.