Occlusion-and-Edge-Aware Depth Estimation From Stereo Images for Synthetic Refocusing

Due to the emergence of dual cameras for smart phones in recent years, synthetic refocusing using stereo data has become an important issue. For synthetic refocusing to produce satisfactory results, it is crucial to render the image with high-quality depth map, which is often obtained through a refinement process. However, most existing depth map refinement algorithms pay little attention to problems related to notorious occlusion and loss of image details. In this paper, we study how the quality of depth map affects the performance of synthetic refocusing and, based on the study, propose a new method that integrates depth information and RGB image for synthetic refocusing. A notable feature of our approach is that it formulates occlusion filling as a labeling problem and solves it by multi-layer alpha matting, resulting in a depth map with edges well-aligned with the RGB image in the occluded area and giving synthetic refocusing a realistic sensation. An evaluation of our method against previous methods is performed by comparing the estimated depth map and the refocusing results.