Research On Light Field Depth Estimation Based On EPI

Compared with ordinary two-dimensional images,light field images can provide the direction information of the light,therefore,the depth estimation based on light field is quite plastic.In recent years,with the advent of light field cameras,more reliable depth estimation on light field has been a hot topic.Occlusion and textureless areas are always challenging problems in light field depth estimation,and little research has been carried out in terms of this matter.To solve these two problems,we proposed a depth estimation method based on a height-adjustable spinning parallelogram operator,which can handle occlusion and textureless areas explicitly.The epipolar plane image(EPI),which is generated by light field images,is composed of many polar lines whose slope is related to the depth of the spatial point.In this paper,we proposed a height-adjustable spinning parallelogram operator and the corresponding distance metric which is used in the EPI.The local depth is estimated by finding the maximum response of the distance metric of all the orientations of the parallelogram to the direction.A common assumption of occlusion is that the occlusion has different colors from the object of interest.Therefore,the polar line of the spatial point in the EPI is truncated,when it is occluded in some sub-aperture images.For the traditional spinning parallelogram operator,only if the pixels of the occlusion are distributed on both sides of the center line,this would lead to errors in depth estimation.In order to reduce the influence of occlusion,we dynamically changes the height of the parallelogram by comparing the similarity between the center point and the points in the center line of the parallelogram window to exclude occlusion.In addition,we proposed a color-consistency cue of the reserved center line,and combined it with pixel distribution distances between the two regions of the parallelogram to form a new distance metric,so as to improve the local depth estimation accuracy of the weak texture region.Finally,the local depth map is optimized by textureless and edge enhancement to get a final global depth map which is smooth,retains more edge details,and has less noise.Experimental results show that the proposed method can handle various occlusion,be robust to textrueless areas,perform well in various complex scenes,and achieve high-accuracy depth estimation.