W循环多重网格的图像立体匹配优化算法

Optimized Algorithm for Stereo Matching Based on W-Cycle Multi-Grid

由于传统的区域立体匹配方法均基于灰度特性,对于纹理少、光照不均匀等情况难以找到正确的对应点。本文提出基于两层W循环多重网格的变分方法对左右图像间进行立体匹配,通过灰度值,梯度值和平滑项这三个不同参数的组合来定义匹配代价函数,并通过实际的立体匹配图像对给予不同的权重值,再通过两层W循环多重网格对匹配代价函数进行优化求得最优解,进一步提高了生成视差图的连续性。通过与真实视差图之间的平方根误差和整体误差两项指标进行对比,结果表明本文视差图出错率比传统立体匹配方法降低20%左右。针对窄基线下纹理特征不明显的重建表面,文中方法鲁棒性较强,可以得到更为平滑准确的视差图。通过反向视差图对求得的牙齿视差图进行对比得到视差分数图,验证了视差图的正确性。

It is difficult to find the correct corresponding points in the case of less texture and uneven illumination by traditional region stereo matching methods which are all based on the gray characteristics.In this paper,a method based on the two-layer W-cycle multi-grid is proposed for stereo matching between left and right images.The matching cost function is defined by the combination of three parameters:gray value,gradient value and smoothing term.Different weight values are given through the actual stereo matching image pairs,and then through the two-layer W loop multi-grid the matching cost function is optimized to obtain the optimal solution,which further improves the continuity of the generated disparity map.The comparison of the square root error and the overall error between the traditional local stereo matching and the new algorithm shows that the error rate of parallax graph is reduced by about 20%.For the reconstruction surface with little texture under the narrow baseline,the new method is so robust that a smoother and more accurate disparity map can be obtained.Finally,the visual fractional map is obtained by comparing the obtained parallax maps with the inverse parallax map,and the correctness of the disparity map is verified.