Depth map filtering method for shape-from-focus recovery based on hybrid network model

Shape-from-focus (SFF) is an economically efficient 3D shape recovery technology. As a core module of 3D digital microscopes, its primary goal is to acquire high-quality depth maps. However, the performance of such technology is highly dependent on the richness of image textures, which inevitably introduces noise into the generated depth maps and seriously impairs their quality. This has become a critical technical challenge in the field. To address this problem, this paper proposes a depth map filtering framework based on a hybrid network model. First, a depth reconstruction method combining the full width at half maximum (FWHM) and Gaussian fitting is designed to generate the initial depth map. Second, a hybrid network model for focus signal classification is constructed, which extracts the differential features of different signals to identify focus signals with noise patterns and generate a flag matrix. Finally, the initial depth map is filtered based on this flag matrix to obtain a high-quality depth map. Experimental results demonstrate that the proposed method can effectively distinguish the patterns of focus signals, filter out depth map noise caused by signal distortion, and simultaneously preserve the integrity of structural details in the depth map, exhibiting excellent filtering performance. In addition, the method proposed in this paper can serve as a general SFF framework, which is applicable to 3D shape measurement or high-end instrument design in different fields.