A Novel Approach For Depth From Focus And Cues From Dcfocus

Filming and visualization of3D information originated in more than one century ago.Traditional photography can only convert3D scene to2D pictures of the scene. Moreover,most of the depth of information has been discarded in this process. In order to have a morepowerful way to render digital effects on photos and, ultimately, to digitalize the real world,refocusing digital images or videos is a critical tool for digital photography and imageprocessing. The key of this post-process is to obtain the corresponding depth map of the scene.Although some information is lost in the process of2D to3D conversion, some cues still left,for example, exploiting defocusing bokeh rendered by lens to identify the degree of blurring.The size of the blur kernel corresponding to each pixel in the photo from the recovered depthmap can be determined. Applications of digital image process after obtaining the depth mapinclude automatic scene segmentation, post-exposure refocusing and re-rendering of the scenefrom an alternate viewpoint, etc.Depth from defocus and depth from focus are important methods to tackle the problem.However, both of these two methods have drawbacks like edge bleeding and failure intextureless regions. In this paper, we propose a novel method of depth from defocus based ona single image and a new approach of depth from focus with multiple images to improve thesetwo shortcomings.By employing mean shift segmentation before the step of building Markov random field,the edges of the recovered depth map are guaranteed to align with the edges of the originalimage, solving the edge bleeding problem. Based on the result of mean shift segmentation, weintroduce confidence in depth from defocus with a single image to enhance the reliability ofthe initial depth estimation of each segment. While in depth from focus with multi-image, weanalyze the response of focus measure in the segment, in order to avoid the tradeoff betweensensitivity of noise and spatial resolution, the inherent problem of fixed window focusmeasure analysis method. After the initial estimation of depth, the hierarchical Markovrandom field is built to expand the area to extract depth information according to the structureof the scene. In this way, our experiments show that depth can extract from the texturelessregions to some extent.