| Depth estimation is an important part of 3D modeling research.The obtained depth information has a wide range of applications in the fields of urban planning,resource survey,rescue and disaster mitigation.Compared with traditional cameras,light field cameras can collect energy information and spatial information of light,including obtaining the spatial position and propagation direction of light.Image refocusing and depth information acquisition can be completed through later digital processing.Since the light angle information acquired by the light field camera occupies the original spatial position information,the spatial resolution of the image is reduced,and large errors are likely to occur in the depth information extraction of pixel matching cues.Therefore,it is of great significance to carry out research on depth information extraction algorithms.In this paper,based on the characteristics of the light field refocusing image,a depth matching algorithm based on pixel color and image gradient cues is proposed.The initial depth map and the center image are used to detect occlusion edges.The joint guidance filter is used to optimize the depth matching cost information.Filter out obvious error points and smooth the deep connection,while retaining the characteristics of blocking the edge.Finally,based on the optimized depth matching data,occlusion cues and confidence,an improved Markov random field model is used to extract the global optimized depth map.The main research contents and conclusions are as follows:(1)Contrast analysis and optimization of refocusing interpolation algorithm: In order to study the effect of interpolation accuracy on the refocusing image during the refocusing process,the images obtained by the refocusing in the spatial domain are first adopted by the neighboring interpolation method,bilinear interpolation method and cubic interpolation method.According to the accuracy and calculation efficiency,the accuracy of the refocused image obtained by cubic interpolation is the highest and the calculation efficiency is the lowest.The refocusing in the frequency domain is optimized based on Sinc interpolation.The experiment proves that the refocused image obtained by Sinc interpolation is compared with the above three interpolation methods.The accuracy is higher;when the same light field image is acquired with different depth refocusing images,the same interpolation algorithm is used,and the frequency domain refocusing efficiency is much higher than the space domain refocusing efficiency.(2)Research on the depth matching clues of the refocused image: According to the consistency of the angular domain color of the refocused pixel point and the consistency of the adjacent area of the spatial domain and the central image,an adaptive pixel color consistent depth matching clue is proposed;according to the gradient of the focus point The consistency of the information and the central image,the minimum matching value of the neighboring area is the focus point matching cost,and according to the spatial domain pixel color consistency constraints,an adaptive image gradient information consistent depth matching clue is proposed;an initial depth map based on the graph cut method is developed Extraction experiments,the results show that both matching clues can better retain the occlusion edge features and reduce the impact of noise.(3)Research on local optimization and global optimization of depth matching cues: In order to reduce the influence of noise,occlusion,texture and other factors on depth estimation,according to the results of the initial depth map and the weighted bilateral filtering of the central image,Canny edge extraction is used for occlusion edge detection;Using the occlusion area mask,the local optimization results of the matching cues based on the guided filtering and the anisotropic diffusion guided filtering are fused;through the optimization of the fracture characteristics of the non-occluded edge depth value,the confidence is estimated to establish the occlusion cues,texture cues and The improved Markov random field global optimization model with confidence level,the calculated final depth map has high overall accuracy,no obvious noise points,and complete occlusion edge information. |
