基于邻域扩展的半自动2D转3D方法*

半自动2D转3D是解决当前3D影视内容匮乏的重要途径,其核心是将用户分配的稀疏深度转换成稠密深度。现有方法大多借助局部邻域进行深度插值,忽略了图像的全局约束关系,因而难以准确恢复深度图的对象边界。针对该问题,提出邻域扩展的最优化深度插值方法。首先引入邻域的邻域,建立邻域扩展的最优化深度插值能量模型；其次在相似的像素点与其邻域加权深度平均值的差异近似相等的假设条件下,将深度插值能量模型的最优化问题转换成一个稀疏线性方程组的求解问题。实验结果表明,与当前流行的半自动2D转3D方法相比,本文方法估计的深度图PSNR更高,同时增强了深度图的对象边界质量。

Semi-Automatic 2D-to-3D Conversion Based on Neighbors Extension

Semi-automatic 2D-to-3D conversion is a promising solution to 3D video creation. One of the key technologies is sparse-to-dense depth interpolation. Existing methods obtain dense depth-map based on similarities in local neighbors which neglects the global constraints in image. Therefore, the object boundaries of depth-map cannot be recovered accurately by using these methods. To help solve this problem, a depth interpolation method using optimization based on neighbors extension is proposed. Firstly, an energy model for depth interpolation is developed by considering distant neighbors. Secondly, we obtain a closed form solution of this model by assuming that similar pixels have similar depth differences between their weighted average value in local neighbors. Experimental comparisons with the popular techniques show that our method demonstrates advantages over PSNR and depth object boundaries.