摘要
基于PatchMatch(同时估计像素点的视差和法向量的3D标签)的方案已经在立体匹配中取得高精度的亚像素视差,但该类方法无法有效解决图像无纹理区域的错误匹配。针对这一问题,对LocalExp(local expansion move)算法进行了改进,并提出一种融合多维信息的自适应像素类别优化的立体匹配算法。该方法设计了一种交叉窗口,在窗口内基于颜色与颜色的自相关信息构建相关权重,并利用约束函数剔除匹配代价中的离群值;在PatchMatch的标签初始化阶段增加约束机制,改进视差标签的建议生成机制,并利用基于局部扩张运动的优化方法求解标签值;利用基于像素类别的填充策略进行视差优化。实验结果表明所提算法能够在Middlebury数据集上取得较低的匹配误差。
PatchMatch-based algorithms that simultaneously estimate the disparities and normal unit of a disparity plane have achieved highly accurate sub-pixel disparities in the stereo matching problem;however, this kind of methods can not effectively deal with error matching in the non-texture regions of image. To solve this problem, we improve the LocalExp(local expansion move)algorithm and present a new stereo matching algorithm integrating multidimensional information for adaptive pixel category optimization. First, a crossover window is designed,the color and color self-correlation information in the window are used to establish the weight, and the restrained function is utilized to eliminate the outliers in the matching cost. Second,the constraint mechanism is added to the label initialization procedure, the proposal generation mechanism is modified, and the local expansion movement algorithm is used to optimize the label values. Finally, the pixel category information-based filling strategy is used to refine the disparity. The experimental results show that the proposed method can obtain a low matching error on the Middlebury dataset.
作者
高雅昆
刘涛
李海滨
张文明
Gao Yakun;Liu Tao;Li Haibin;Zhang Wenming(Key Laboratory of Indust rial Computer Control Engineering of Hebei Province,Yanshan University,Qinhuayigdao,Hebei 066004,China)
出处
《光学学报》
EI
CAS
CSCD
北大核心
2019年第7期261-273,共13页
Acta Optica Sinica
