Research Of Depth Recovery From Defocus Analysis

Recovering depth information of3D object from2D images is an importanttask of computer vision,and also is the most studied, the most widely used andfastest growing area. Depth information recovery technique can be widely appliedin3D reconstruction, robot navigation, medical imaging, biometrics informationextraction, industrial testing and other fields. In recent years, due to thetraditional multi-vision depth recovery algorithm computational complexity andeasy to produce wrong matching, monocular vision contrast with it has its uniquecharacteristics there for attracted much attention.Defocus information is the mostimportant clues used in monocular visual depth recovery.This paper starting from the physical principle of defocus fuzzy model, usingthe basis of the relationship between the point spread function and depthinformation, combined image processing method with defocus fuzzy theory,discussion the defocus information in the application of depth recovery. Includingsingle defocus image and multiple defocus images depth recovery. The main workof this paper are as follows:(1) The problem of depth recovery of a scene from a collection of defocusedimages.When the model of the PSF is known, we propose an optimal method toinfer depth information from defocused images that involves computing orthogonaloperators which are regularized via functional singular value decomposition. Whenthe model of the PSF is unknown, we propose a simple and efficient method thatfirst learns a set of projection operators from blurred images and then uses theseoperators to estimate the depth information of the scene from novel blurredimages.Our method overcome the disadvantages of high computational cost andweek applicability in traditional method.(2)In this paper, we also address a challenging problem of depth informationfrom a single photograph taken by an uncalibrated conventional camera. In order toachieve this, a rough blur map of the input image, which reflects the amount ofdefocus blur at edge locations, is obtained by local robust blur estimation, then theimage segmentation method is applied to propagate the blur value from edge locations into the unknown regions,then recovery the scene depth information. Weshow that our method is robust to noise, inaccurate edge location and interferences ofneighboring edges and can generate more accurate scene depth maps compared withexisting methods. We also discuss ambiguities arising in recovering depth from singleimages using defocus cue and propose some possible ways to remove theambiguities.Experimental results demonstrate the effectiveness of our method in providinga reliable estimation of the depth of a scene.