Scale Space Based Light Field Local Depth Estimation With Global Optimization And System Implementation

Light field photography,as a novel way to describe the visual appearance of a scene,has been rapidly evolving in recent years.By placing a micro-lens array in front of an image sensor,lenslet-based light field camera can collect light rays coming from different directions,which allows users to alter the point of view or focal length after a single shot.The abundant information light field carries benefits some traditional computer vision tasks,including depth estimation.At the same time,the special structure of the light field image also poses new challenges for depth estimation.In this thesis,we work on the dense depth estimation of light field images,and mainly completes the following work:1)A phase correlation matching algorithm based on scale space is proposed to estimate the depth of the light field image.In this thesis,the imaging principle and characteristics of the light field image are analyzed.In view of the micro-baseline of the light field,the phase correlation algorithm is applied to the stereo matching of the light field image in sub-pixel accuracy,and the scale space is constructed by the way of the light field image pyramid,which improves the matching precision and efficiency.2)Based on the phase correlation metric,this thesis implements a light field occlusion edge detection algorithm for improved depth estimation.Based on the phase correlation translation-invariant root mean square error,the occlusion edges of the light field image are extracted.Combining the occlusion model of the light field,this thesis adaptively select the target viewpoint in depth estimation process.For the occlusion edge,the depth estimation is performed based on the refocusing method,which avoids the error caused by the local window in the stereo matching.3)In the global optimization stage of depth estimation,the energy function is designed via considering the occlusion in the form of weights.The Markov random field energy minimization optimization method is used to improve the results.In some optimization methods such as guided filtering,the edge of the non-occlusion area may be strengthened in the depth map.This thesis explicitly considers the occlusion in the energy f-unction,so as to achieve the effect of smoothing the interior of the obj ect and emphasizing the edge of the object in the depth map.In summary,within a multi-resolution framework,this thesis extracts image signals by a local window for phase correlation matching.At the same time,the occlusion edges are detected,and special processing is performed on the occlusion edges.Estimation is guided layer by layer between multiple resolutions to obtain a dense depth map of the light field image.Experiments show that the proposed algorithm achieves the sub-pixel precision depth map with high efficiency.