摘要
尽管传统的立体匹配模型在精度和鲁棒性方面都表现出了良好的性能,但在弱纹理和深度不连续区域的视差精度问题依然存在。针对上述问题,提出了一种基于改进匹配代价和均值分割的最小生成树立体匹配算法。首先,在匹配代价计算阶段,通过Census变换进行初始匹配代价计算,利用Sobel算子对输入图像进行边缘信息提取,将提取后的图像边缘信息与Census变换后的匹配代价值进行融合,并将其与基于图像亮度信息的代价值进行非线性融合,以提高匹配代价的精度;然后,使用最小生成树代价聚合模型进行聚合操作并利用赢者通吃策略估计图像的初始视差;最后,在视差优化阶段,采用MeanShift算法对图像进行分割,结合图像的轮廓信息对误匹配点进行修正,进一步提高在弱纹理及边缘区域的视差精度。实验结果表明,与一些传统算法相比,所提算法具有更高的视差精度,且视差图的边缘、纹理较其他算法更为平滑,具有更强的鲁棒性。
The disparity accuracy issue still exists in weak texture and depth discontinuity areas,although the conventional stereo matching model has demonstrated good performance in accuracy and robustness.To address the above issues,a minimum generating cube matching algorithm based on enhanced matching cost and mean segmentation is proposed.First,in the matching cost computation stage,the initial matching cost is computed by the Census to transform,and the input image’s edge information is extracted by the Sobel operator.The extracted image edge information is merged with the matching cost value after Census transform,and the nonlinear fusion is conducted with the cost value based on image brightness information to enhance the matching cost’s accuracy.Then,the minimum spanning tree cost aggregation model is employed for aggregation operation and the winnertakeall technique is employed to estimate the image’s initial parallax.Finally,in the disparity optimization stage,the MeanShift algorithm is employed to segment the image,and the mismatching points are corrected along with the image’s contour information to further enhance the disparity accuracy in weak texture and edge areas.Experimental findings demonstrate that compared with some conventional algorithms,the proposed approach has higher disparity accuracy,and the disparity map’s edges and textures are smoother and more robust than other algorithms.
作者
王道累
韩洋
Wang Daolei;Han Yang(College of Energy and Mechanical Engineering,Shanghai University of Electric Power,Shanghai 200090,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2023年第4期323-332,共10页
Laser & Optoelectronics Progress
基金
国家自然科学基金(61502297)。
