Abstract:

Because the micro lenses segment the plenoptic image physically, the global optimization method is hard to be applied on the depth estimation progress and the WTA (winner-take-all) method based on the multibaseline system will generate very coarse but continuous depth map. The global optimizations method always generate robust but concrete depth map when using multi-label technique. So cost projection strategy by directly processing the lenslet plenoptic image is proposed to build the cost volume based on the projection image for global optimization. A method based on Markov Random Fields Propagation (MRFsP) is adopted by merging the modified WTA depth result to refine the concrete depth map. For validation, the depth map built with the proposed method is compared with the microlens-based optimization result, and the result shows obvious improvement.

Key words: plenoptic image, depth estimation, Markov Random Fields Propagation (MRFsP)

摘要：

由于微透镜影像分割了原始全光影像, 无法应用全局优化算法, 而采用WTA (winner-take-all)算法得到的深度影像虽然连续, 但包含错误估计的深度信息; 利用全局优化计算得到的深度图可以消除错误估计的深度信息, 但是受到代价计算时深度离散化的限制而不连续。基于上述情况, 提出一种直接处理微透镜阵列全光影像的代价投影策略, 在投影图像上构建代价立方体(cost volume, CV), 同时采用MRFsP (Markov Random Fields Propagation)进行优化, 结合WTA方法的优势改善离散深度图像。实验结果表明, 与现有方法对比, 利用MRFsP优化后的深度图既能消除错误匹配点, 也能保持连续。

关键词: 全光影像, 深度估计, MRFsP优化