Depth recovery from a single defocused image based on depth locally consistency

Demand to depth estimation from a single image is emerging in more and more multimedia applications, such as human-computer interaction, 3D video generation and representation. In this paper, we manage to solve this challenging problem by using the defocused cues contained in the image. Our algorithm is based on depth local consistency assumption that the scene in input image can be modeled as multiple planar surfaces, and each over-segmented patch corresponds to a small p lanar surface in the 3D scene. As a result, a graph-based algorithm is firstly applied to obtain an over-segmentation result of the input image. Based on the sharp edge prior, the blur amounts on the boundaries of segmentations are calculated according to the edge width at the corresponding locations. Then, the blur amounts of the unknown regions are interpolated from the boundary regions using a plane fitting method. In the next, an image guided filtering method is applied to refine the obtained blur map. To eliminate the affect of the tiny textures, an L0 gradient minimization algorithm is applied to the input image to preserving the prominent boundaries, and the resulting image is served as the guided image. At last, based on a simple geometry prior of photograph, a binary graph cut algorithm is adopted to eliminate the ambiguity in the depth map over the focal plane. The performance of our algorithm is evaluated by various test images. The results demonstrate our algorithm produces high quality depth maps, outperforming state of the art.