Research On Depth Estimation From A Single Defocused Image

In the research field of Computer Vision,3-dimensional image reconstruction technique has been a hot issue.And acquiring the object's 3D information directly is limited by expensive hardware devices,so using two-dimensional image recovering depth information of the target scene is studied.The depth of the objects in the two-dimensional image refers to the distance between the object to the camera lens in the process of filming.Traditional depth recovery methods based on the double(multiple)visual can restore the target scene's high precision depth information,and have been widely studied and used.Due to their high computational complexity and wrong matching easily,the monocular vision depth recovery methods with advantages like simple computing and good real time ability attracted more and more attention.One of the most important clues in the monocular visual depth recovery is defocus information.Depth from defocus(DFD)is the representative of the monocular visual depth recovery method.It can not only effectively avoid numerous images of the target scene and complex calculation process,but also has the advantages of high real time capability.DFD estimates the depth of the object through comparing the difference between defocus degrees in the same scene,so quires a pair of images of the same scene with different defocus setting are needed.The constraints of secondary imaging may increase the operation complexity of the algorithm,which results in some limits in real application.To overcome the shortcomings of traditional DFD,this thesis carries out researches on single defocused image depth recovery method based on monocular defocus blur clues.The main works are as follows:(1)Depth estimation from a single defocused image based on Gaussian-Cauchy mixed model.Most methods for depth estimation from a single defocused image construct the point spread function by an 2D Gaussian or Cauchy distribution.However,reasons of blurred images in the real world are varied,so a simple Gaussian or Cauchy distribution function may be not the best approximation model.They are often influenced by noise and inaccurate edge location,and then a high quality depth estimation may be difficult to achieve.This thesis presents a Gaussian-Cauchy mixed distribution model to re-blur the given defocused image,and two different degree blurred images are then obtained.The sparse depth map generated from the gradients ratio at edge locations by the two blurred images is estimated.In so doing,a full depth map can be recovered by matting Laplacian interpolation.Experimental results on some real images demonstrate that the proposed approach is effective and better than the two commonly used approaches.(2)Depth estimation from a single defocused image based on super pixel segmentation.Existing DFD methods generally compute the blur at edge locations and solve an optimization problem to propagate the blur from edges to all image pixels.Solving the pixel-based optimization problem is time-consuming,posing the performance bottleneck.Moreover,the generated depth maps are not consistent in textured areas and the blur estimation may be incorrect in the regions with soft shadows.These problems are solved by proposing a superpixel-based blur estimation method.Experimental results show that the proposed method is not only faster than pixel-based blur estimation,but can improve depth data in textured regions and soft shadows as well.The proposed two depth estimation methods for a single defocused image may improve the recovery effect and efficiency.The obtained results will enrich the research of using two-dimensional image recovering depth information of the target scene.