图像引导的二阶总广义变分稀疏深度图的稠密重构

利用图像颜色信息进行深度图重构,可以恢复对象边界处的深度不连续性,但无法保证对象内部的深度均匀性。为解决该问题,提出图像引导下总广义变分正则化的深度图重构模型。该模型利用扩散张量将图像提供的边缘信息引入二阶总广义变分正则项,使得重构深度在保持对象边缘的同时逼近分段仿射平面,从而保证恢复深度既保持对象边界处的不连续性,又具有对象内部的均匀性。通过Legendre-Fenchel变换将模型转换成等效的凸凹鞍点问题,从而得到高效的一阶原始对偶求解算法。实验结果表明,该方法能够恢复尖锐的对象边缘,同时保持对象内部的深度均匀性。与现有算法相比,所提方法具有更高的峰值信噪比、归一化互协方差和更低的平均绝对误差。

Dense Depth Map Reconstruction via Image Guided Second-order Total Generalized Variation

The depth map reconstruction using image colors may recovery the depth discontinuities at object boundaries,but will damage depth uniformities inside objects.In order to solve this problem,we formulated depth reconstruction as a convex optimization problem which is regularized by image guided total generalized variation.By incorporating image diffusion tensor into the variation regularizer,the proposed method generates piecewise smooth depth while preserving discontinuities at object boundaries.To efficiently solve the problem,a first-order primal-dual scheme was derived based on the Legendre-Fenchel transformation.Experimental results demonstrate that our method can preserve depth discontinuities at object boundaries and uniformities inside objects and outperform existing methods in terms of peak signal-to-noise ratio,normalized cross-covariance and mean absolute error.