Robust Light Field Depth Estimation over Occluded and Specular Regions

Traditional methods for light field depth estimation establish the cost data to measure the photo consistency of pixels refocused into a specific depth range, with the highest level of consistency indicating the correct depth. These methods are based on the photo consistency of Lambertian surface. However, the photo consistency is broken when occlusion and specular reflection occur. In this paper, a new depth estimation algorithm is proposed to solve the problem that the photo consistency is broken. Firstly, the central view image is segmented into multiple superpixel regions. The cost ranges of the un-occluded points and occluded points in the refocusing process are analyzed, and a penalty term is added to the pixel whose color deviation exceeds an adaptive threshold to detect the occluded points. Because the un-occluded pixels in the angular sampling image still keeps the photo consistency, we propose a voting method to select the un-occluded pixels to obtain the initial depth of the occluded point. We use a method to determine the specular region based on similar features of color and texture in the superpixel region and then present an optimization energy function to obtain the depth of the specular region. Finally, a more accurate depth map is obtained by using a globally optimization. Experimental results show that the proposed method is superior to other comparison algorithms, especially in the cases of the specular regions and multi-occlusion.