医用光学中为快速消除双目立体匹配歧义,提出一种基于局部特征、稳健高效的两步立体匹配方法。用局部二进制模板/对比度(LBP/C)纹理分析描述图像纹理并构造初始匹配成本；仅在垂直、水平方向内自适应地分配权值,消除匹配特征的相似歧义性,并行双通地聚合匹配成本；由赢者通吃法得到初始视差。为消除弱纹理、重复纹理和遮挡等引起的歧义,视差求精方案包括并行双通校准视差、基于映射检测遮挡、基于极线最小二乘拟合填充遮挡。实验表明,该算法计算效率高、结构简单,易于实现并消除匹配歧义,得到精度较高且分段平滑的视差。

A feature-based two-step local stereo matching with high efficiency is presented to eliminate the ambiguity of binocular stereo problem in medical optics. The local binary pattern/constrast (LBP/C) textures describing local features are firstly analyzed for both stereo pairs,and the initial matching costs is constructed from the features. Two-pass aggregation with adaptive support weight (ASW) for the costs are adopted only in the vertical and horizontal directions,which efficiently resolves ambiguity in feature matching. Initial disparity map is obtained via local Winner-Takes-All optimization from them. To solve the ambiguity from other aspects including textureless,repetitive patterns and occlusion etc.,disparity refining procedure consists sequentially of three steps:two-pass ASW-based disparity calibration,warping-based occlusion detection and epipolar-line-based least square fitting to handle occlusion. The experimental results indicate that this approach with texture cues and low complexity has high efficiency and concise structure,being easily implemented and eliminting ambiguity,so that it can obtain comparably accurate and piecewise smooth dense disparity map effectively.