Attention based dual path fusion networks for multi-focus image

Nana Yu, Jinjiang Li, Zhen Hua

Abstract: An important goal of multi-focus image fusion technology is to generate all-focus images that can better retain the source image information, while improving the quality and performance of image fusion. However, traditional image fusion methods usually have problems such as block artifacts, artificial edges, halo effects, and decreased contrast. To solve these problems, this paper proposes a dual-path fusion network (A-DPFN) with attention mechanism for multi-focus image fusion. Firstly, our method splits the complete image into image blocks, and obtains higher image classification with the preprocessing of the image blocks, so that our proposed dual-path fusion network accelerates the model convergence speed; Secondly, feature extraction block1 (FEB1) and feature extraction block2 (FEB2) in our network respectively extract the feature information of pair of focused images, in which we have added an attention mechanism; Finally, we merge the obtained pair of feature images as the input of the feature fusion block (FFB), and enhance the details of the image through the down-sampling block (DB) and the up-sampling block (UB). The experimental results show that the method has strong robustness and can effectively avoid problems such as block effect and artificial effect. Compared with the traditional image fusion method, the multi-focus image fusion method proposed in this paper is more effective.