重新下降M估计子约束的深度图超分辨率算法

Depth-map Super-resolution Algorithm with Redescending M-estimator Constraints

现有的深度图超分辨率增强算法大多借助于同场景彩色图像提供的辅助信息,而不同传感器信号间的结构差异将会引入质量损伤.为此,将图像引导的深度近邻关系视为误差,并利用重新下降M估计子进行误差的测度,从而有效抑制彩色图像和深度图像间结构差异的问题.首先根据相似颜色具有相似深度的假设建立深度近邻约束;其次利用重新下降M估计子度量深度邻域约束,将深度超分辨率增强转换成一个最优化问题;最后通过广义迭代重新加权最小二乘法予以求解.实验结果表明,该算法可有效地保持深度图的对象边缘,定性和定量指标均优于现有的代表性算法.

Most of the depth-map super-resolution algorithms rely on the information provided by the guidedcolor image. Due to differences in structure between guided and input signals, such algorithms are hard to preservedepth boundaries. We address this problem by redescending m-estimators. First, the neighboring constraintsfor depth are built based on color similarities. Second, redescending m-estimator is used to measure the constraints.Then, the depth super-resolution is formulated as an optimization problem. Such a choice helps in dealingwith violations of the assumption that similar colors have similar depth. Finally, the solution is obtained by thegeneralized iteratively reweighted least squares. The experimental results demonstrate that our algorithm can preservedepth boundaries and is superior to existing algorithms in terms of depth accuracy.