融合一致性与差异性约束的光场深度估计

Light field depth estimation of fusing consistency and difference constraints

光场图像深度估计是光场三维重建、目标检测、跟踪等应用中十分关键的技术。虽然光场图像的重聚焦特性为深度估计提供了非常有用的信息,但是在处理遮挡区域、边缘区域、噪声干扰等情况时,光场图像深度估计仍然存在很大的挑战。因此,提出了一种基于极平面图(Epipolar plane image,EPI)斜线像素一致性和极平面图区域差异性的深度估计算法用于解决遮挡和噪声问题。EPI斜线像素的一致性采用旋转线性算子(Spinning linear operator,SLO)的颜色熵度量,能够提高深度图边缘的准确性以及抗噪能力;EPI区域的差异性采用旋转平行四边形算子(Spinning parallelogram operator,SPO)的卡方χ^(2)度量,能够提高深度图深度渐变区域的准确性,并使用置信度加权的方法将两种度量进行融合,可以减少遮挡区域和噪声的干扰。另外,充分利用像素邻域的颜色相似性,使用引导保边滤波器和马尔科夫随机场(Markov random field,MRF)全局优化策略进行后处理,进一步减少深度图的边缘错误,得到遮挡边缘准确的深度图。在HCI光场数据集上进行了实验,并与经典光场深度估计算法进行了对比,结果表明该算法在主观质量和客观指标两方面都有明显提升。

Light field image depth estimation is a critical technique in applications such as light field 3 D reconstruction,target detection and tracking.The refocusing property of light field image provided rich information for depth estimation,however it was still challenging in the case of occlusion region,edge region,noise interference,etc.Therefore,a depth estimation algorithm based on the consistency of epipolar plane image(EPI) slant pixels and the difference of epipolar plane image regions was proposed to solve the occlusion and noise problems.The consistency of EPI slant pixels was adopted by the spinning linear operator(SLO) color entropy metric,which could improve the accuracy of depth map edges as well as the noise immunity;the difference of EPI regions was measured by the chi-square χ^(2) metric of the spinning parallelogram operator(SPO),which could improve the accuracy of the depth gradient area of the depth map,and the two metrics were fused with confidence weighting,which could reduce the interference of occluded regions and noise.In addition,the color similarity of the pixel neighborhood was fully utilized,and post-processing was performed with a guided edge-preserving filter and Markov random field(MRF) global optimization strategy to further reduce the edge error of the depth map to obtain an accurate depth map with occluded edges.Experiments were conducted on the HCI light field data set and compared with the classic light field depth estimation algorithm.The results show that the algorithm has a significant improvement in both subjective quality and objective indicators.;that the proposed method outperformed state-of-the-art depth estimation methods.Compared with the classical light field depth estimation algorithm,proposed method had significant improvement on both subjective and objective qualities.