Multi-scale siamese networks for multi-focus image fusion

In this paper, we propose a multi-scale Siamese network for multi-focus image fusion. Many current image fusion methods are based on classifier and discriminators to segment the original image, determine whether there is a focus on it, and generate the fused image by post-processing the decision map. We input two complementary source images as two branches into the network, introduce multi-scale convolution module and pyramid attention to extract image information, the whole process of fusing image edge information of different scales, reduce the information loss that occurs in the fusion process, and can better deal with the problems of artifacts and small-scale visual blurring of the images obtained by existing fusion methods, and the fused images are more adapted to the human visual. The proposed method is quantitatively and qualitatively compared with other existing more advanced methods on three publicly available datasets to demonstrate the effectiveness and superiority of the performance of our proposed method.