基于跨尺度PatchMatch的立体匹配算法

针对现有的PatchMatch（3D标签优化）立体匹配算法存在对图像中弱纹理、视差不连续区域匹配精度低的问题，提出了一种结合超像素分割和跨尺度PatchMatch的立体匹配算法。首先，通过高斯下采样获得多尺度图像并对各尺度图像超像素分割。其次，基于四色定理腐蚀超像素边界使3D标签在超像素上迭代传播具有子模性和独立性，生成的子模能量用图割（Graph Cut，GC）算法得到最优解。最后，提出跨尺度能量函数模型，约束不同尺度下同名像素3D标签能量一致，使3D标签迭代传播可在不同尺度进行GC优化，获得最优视差图。在Middlebury数据集上的实验结果表明，本文算法对21组弱纹理、复杂纹理图像的平均误匹配率为2.20%，相比其他改进的PatchMatch立体匹配算法误匹配率降低了10.1%，且视差图误匹配可视化显示，弱纹理、视差不连续区域匹配效果优于其他改进的PatchMatch立体匹配算法。

Stereo matching algorithm based on across-scale PatchMatch

To solve the problem of Insufficient matching accuracy of the weak texture and disparity discontinuity regions in the image, A stereo matching algorithm combining superpixel segmentation and cross-scale PatchMatch is proposed. Firstly, multi-scale images are obtained by the gaussian under-sampling, and superpixel segmentation of each scale image. Then, based on the four-color theorem, corroding superpixel boundaries makes the iterative propagation of 3D labels on superpixels sub-modular and independent, and the generated sub-modular energy is optimized by the Graph Cut (GC) algorithm. Finally, aim to make 3D label iterative propagation can be cross-scale GC optimization to obtain the optimal disparity map, a cross-scale energy function model is proposed to constrain the consistent energy of 3D labels of the same pixel at different scales. Experimental results on Middlebury data set show that the average mismatch rate of the proposed algorithm for 21 groups of weak texture and complex texture images is 2.20%. Compared with other improved PatchMatch stereo matching algorithm, the false matching rate is reduced by 10.1%. Visualization of disparity map mismatched regions shows the proposed algorithm is better than other improved PatchMatch stereo matching for weak texture and disparity discontinuity regions algorithm.