Depth Estimation And Application From Stereo Vision And Spectrally-varying Depth-of-Field

Estimating the depth information of the scene from the two-dimensional image is one of the classic problems in the field of computer vision.The main purpose is to obtain the distance between the scene and the camera,which is available to realize tasks like three-dimensional reconstruction of the scene,automatic scene segmentation,post-exposure re-focusing and scene re-rendering under different point of view.Binocular stereoscopic vision is one of classical methods of depth estimation,which matches corresponding points in left-right rectified image pairs and calculates depth information from the difference of coordinates of corresponding match points.However,issues often occur such as occlusion and mismatching due to the different locations of each point of view,resulting black holes and discontinuity in depth map.Another traditional method to estimate depth information is called as depth from defocus(DFD).The principle is based on the deviation degree of the fuzzy degree of defocused image formation,which is related to the distance between object and image plane.The greater the deviation distance,the greater the degree of fuzzy.Unlike stereo vision,the image required in this method is obtained by changing the camera parameters with a steady view point.The algorithm is usually based on the image texture gradient as focus criterion,so it is suitable for the depth estimation of the area with large image gradient.Due to the solution of point spread function(PSF)of each pixel point is a ill-posed problem,an approximate solution is often used in practice,causing low precision and poor robustness of this algorithm.The main contributions of this paper include two aspects.First,the traditional method of defocus is improved,and a coded lens is designed.By modifying the aperture of the conventional color camera,the effective aperture size of a color channel is smaller than that of the other two.Secondly,in view of the advantages and disadvantages of the two traditional methods,In this paper,a depth estimation method combining binocular parallax method and multi-spectral variation is proposed,which solves the problem that the binocular parallax method and the multi-spectrum variation fusion data set are obtained and the robustness of the single method is low for the discontinuous region.An algorithm combined depth map and image matting is designed.Among them,the depth map provides a certain set of prospects,background sample points for image matting,image matting implies rich boundary information to make up for the depth map in the boundary processing defects.The two are iterated to achieve the effect of automating the matting and enhancing the depth map.Finally,experiments are carried out in the Middlebury dataset and the real scene respectively,which proves the superiority and reliability of the proposed method.