Light field depth estimation based on sub-pixel displacement and multi-directional pixel gradient

Abstract

Depth estimation based on light field imaging techniques has significant application value in industrial defect detection and medical imaging fields. Researchers have continuously developed and optimized algorithms such as epipolar plane images (EPI), multi-view stereo matching, and digital refocusing to achieve high-accuracy depth estimation. In recent years, neural network-based approaches also exhibit excellent performance in synthetic datasets, although these methods still face significant challenges such as efficiency, generalization capability, robustness, and precision for realistic applications. In this paper, we propose a method that combines sub-pixel disparity displacement and multi-directional pixel gradient symmetry to calculate the depth information. Specifically, we employed the stereoscopic interpolation for high-accuracy refocused imaging sequence regeneration, and achieved depth estimation by analyzing the symmetry of pixel gradients. To further optimize the algorithm’s performance, we introduce a four-direction evaluation to handle occlusion scenarios, which greatly improves confidence results from $87.05\%$ @ single direction to $94.17\%$. From comparative experiments, the proposed method obtains more accurate estimation results than the classical conventional algorithms. Meanwhile, it also demonstrates much better generalization capabilities and speed in real world images compared to the recently reported neural network-based approaches.