Study Of Depth Estimation Based On Multiple Cues For Light Filed Cameras

As an important branch of computational imaging,light field imaging technology shows satisfactory development potential when assists in solving some problems of computer vision.Compared with the conventional camera,light field camera could capture the four-dimensional light field of scene through a single exposure,which contains spatial information and angular information.Thus light field camera has the capability of recovering the depth information.However,light filed camera has poor completeness when samples the light filed signal.Affected by the complex scene,depth recovery from a light field camera still faces the challenges on accuracy and robustness,especially at the depth discontinuities.This paper focuses on the presence of occlusions,and presents a depth estimation method using the light field camera which base on some depth cues.This paper first introduces some depth cues in light field data,such as defocus cue,correspondence cue and occlusion cue.According to the characteristics of digitals refocusing,we introduce a method to calculate the defocus cue of focal stack which is based on the discrete cosine transform.According to the relationship with refocusing and angular coherence,and the photo consistency,we show a method to estimate depth using the correspondence cue.Then we introduce two methods to measure the correspondence responses,which using variance and the angular entropy.Meanwhile,this paper discusses the occlusion problem.In the presence of occlusion,the assumption of angular coherence and photo consistency will fail.So we develop an occluded-ray culling method under the 3-sigma rule,and then use the rest rays to compute the correspondence responses.In the next,we continue to discuss the occlusion in light filed.We discuss that the occlusion cue has different manifestations at ray-level,pixel level and image level.On the basis of this consequence,we produce a simple but efficient method to solve the problem of occlusion.We propose a cascade occlusion culling filter to filter or reduce the influence from occlusion information.We conclude that these three occlusion culling filters could,to some extent,improve the accuracy of depth,especially at the depth discontinuities regions and occluded regions.Finally,we propose a depth estimation method based on cascade occlusion culling and multiple depth cues.Based on the three level cascade occlusion culling filter,we combine the defocus cue,correspondence cue and occlusion cue,and translate the depth optimization problem to another problem which minimizes a defined energy function.We use the graph cut optimization framework to solve the minimization.According to the experimental results,our method could get accurate depth at depth discontinuities regions and texture discontinuities regions.