Mutli-focus image fusion based on guided filter and image matting network

The problem of limited depth-of-field is one of the major disadvantages in optical imaging devices, whereas multifocus image fusion(MFIF), as an enhancement technique, can effectively solve this situation. The purpose of MFIF method is to fuse the different focus regions from the multifocus image pair so as to obtain a full-clear fusion result. However, the inaccurate decision maps in fusion methods often leads to blurred object boundary or artifact. To solve this problem, we propose a novel method combining guided filter and deep image matting network for multifocus image fusion tasks. The first step is to utilize guided filter to generate high-frequency information of multifocus image pairs, as to get the initial decision maps based on pixel-level maximum comparison. Aiming for the better fusion images, these decision maps need to be further refined. Then, the trimaps corresponding to the initial decision images are gained via the basic morphological operations, and we utilize the deep image matting network to refine the uncertain regions of trimaps, which can effectively reduce the misclassified pixels and yield the final decision maps with less noise. Finally, the fusion results are constructed by using the final decision maps. This approach effectively distinguishes between focused and non-focused areas, preserving the complete and sharp contour information of objects in the fused images. The experiments illustrate the superior visual effect and objective evaluation on public datasets are achieved by the designed fusion method compared with the advanced fusion methods.