A Denoising Depth Estimation Algorithm Based on 4D Light Fields

With the progress of science and technology and people’s need for cognition about their surroundings, depth estimation has been a highly popular research topic for decades. In recent years, with the rise of market demand and industrial development, the research of depth estimation has even been extended to a variety of applications, such as visual reality, robot navigation, and the defect detection in the industrial production lines. In this thesis, the two dimensional spatial information of a sequence of images by image processing technology is utilized to develop a depth estimation algorithm. The related studies of depth estimation are reviewed, including stereo vision and light field photography, the two main image-based depth estimation methods. This thesis focuses on the light field photography method. The four-dimensional light fields that are acquired from plenoptic camera can record much more light information coming from different directions, but the traditional camera can just contain a part of the light information. Via transformations, a sequence of images called sub-aperture images and epipolar plane images can be obtained. we use the 4D light field data to propose a depth estimation algorithm, combining epipolar plane image analysis and our denoising processing based on Markov Random Fields and LoG filter. In the results, it is shown that our proposal can obtain the consistent depth with the scene with higher accuracy and efficiency.