基于稀疏-密集匹配算法的视差度量

图像匹配是一种约束最优化问题，系统是否收敛于全局最优值一直尚未解决。为得到密集连续的视差图，通常在匹配算法中引入唯一性、连续性、有序性约束。但对于遮挡区域的像素，其对应点并不存在，因此过度平滑的视差结果会导致在遮挡边缘处，灰度变化显著的表面渗透至邻接的灰度变化和缓的表面，出现所谓的表面膨胀缩小现象。当定义图像匹配算法中最优匹配标准时，必须根据图像的内在信息加入其他一些约束条件，保证遮挡存在下匹配结果的全局精确度。提出一种稀疏-密集的匹配算法，利用能量图确定稀疏的视差图，然后运用相位匹配方法对视差图进行内插，并结合最优化理论克服局部最优困扰。实现遮挡检测。立体像对的匹配试验结果表明了该方法的有效性。

Measurement of disparity based on sparse-to-dense matching approach

Image matching belongs to constrained optimization problems. Whether the system would converge to the global optimum is still an open problem. To produce smooth and detailed disparity map, three assumptions--uniqueness, continuity and ordering--are generally adopted. However, if a point appears in one image but it is occluded in the other one, above assumptions are invalid. While the unwanted smoothing occurs in the resultant disparity map, a surface with high intensity variation extends into neighboring surfaces with less variation across occluding boundaries. This fact creates the phenomena of so-called fattening and shrinkage of a surface. Therefore, when considering the definition of criteria for an optimal match in one matching algorithm, some other matching constraints must be imposed based on the internal image information, to ensure the global accurate matching results accompanied with occlusion detection. A sparse-to-dense matching approach is presented, which utilizes the energy map to gain sparse but high confidence disparity map, the phase matching to interpolate the disparity results and the optimization theories to avoid local extrema and detect occlusion areas. The experimental results demonstrate the validity of the proposed approach.