Research Of Algorithm Of Depth Recovery From A Single Defocused Image

To recovery depth information from 2D images is a classic problem in computer vision. The main purpose is to get the distance from each object to the camera for 3D reconstruction and so on. From the principle of geometric projection, we find the non-linear relationship between defocused blur and depth information. Recent years the algorithms to recover depth from defocused images arouse many researchers’attention. With this method depth values can be obtained from the edges’blur values based on the assumption that each edge in the image is sharp. And then get the full blur map by propagating blur amounts from the sparse blur map. Nevertheless, in current algorithms a problem exists that is too much detail in the input image was introduced into the depth map which has influenced the final result. In order to address the problem, in this thesis we apply super-pixels and LO gradient minimization to reduce the introduction of the detail. Additionally, we adopt iteration to refine the depth map based on the relationship of segmentation and depth estimation. In detail, the main points of this thesis would be as follows.(l)Firstly, we propose a new principle-the locally consistency which means each area with consistent attribute in the 2D image correspond to a plane in the 3D scene and its depth values varies a little. According to the assumption, we apply the graph-based segmentation to get the over-segmentation result, which can ensure each segment satisfy the locally consistency.Secondly, in order to avoid more detail being brought into the final result, we introduce another method-LO gradient minimization. In this way, we can reduce tiny texture while the prominent boundaries are preserved, so that the errors caused by the details from the input can be fixed.(2)From the experiments, we find the relationship between image segmentation and depth estimating. We can estimate depth by segmenting image while depth information can also optimize the segmentation. Therefore, the iterative method is introduced into the depth recovery aiming at optimizing segment with depth information. As the terminal condition, we proposed the depth map dissimilarity degree. As all described above, the main purpose is to get better depth map.