摘要
在深度图像处理的研究领域,立体匹配技术长期面临深度不连续、遮挡、弱纹理区域及抗噪声等多重难题。特别是局部立体匹配算法在弱纹理区域的精度较低,且易受到噪声干扰。针对这一问题,提出了一种基于四状态Census变换的改进方法。同时,将梯度信息与优化后的Census变换融合,设计了一种新的代价计算方法,从而提高了匹配精度并增强了抗噪声能力。在代价聚合阶段,引入了多种自适应窗口策略,根据每个窗口的特性设置相应的自适应阈值,从而优化了算法的灵活性与准确性。最后,在视差优化阶段,采用赢家通吃(WTA)算法对初步视差进行精细校正,确保最终视差图的高精度。通过在Middlebury测试平台上对标准立体图像数据集的实验验证,所提算法表现出优异的性能,平均误匹配率仅为5.05%。与传统算法相比,所提方法在匹配精度上实现了提升,并且在复杂环境下展现出了更强的稳健性。
In the field of depth image processing,stereo matching technology has long faced multiple challenges,including depth discontinuities,occlusions,weak-texture regions,and noise susceptibility.In particular,local stereo matching algorithms exhibit low accuracy in weak-texture regions and are easily disturbed by noise.To address this issue,this paper proposes an improved method based on the four-state Census transform.Meanwhile,by combining gradient information with the optimized Census transform,a new cost calculation method is designed to improve matching accuracy and enhance anti-noise capability.In the cost aggregation stage,multiple adaptive window strategies are introduced,with corresponding adaptive thresholds set according to the characteristics of each window,thus optimizing the flexibility and accuracy of the algorithm.In the disparity optimization stage,the Winner-Takes-All(WTA)algorithm is adopted to finely correct the initial disparity,ensuring high accuracy of the final disparity map.Experimental verification on standard stereo image datasets using the Middlebury test platform shows that the proposed algorithm achieves excellent performance,with an average mismatching rate of only 5.05%.Compared with traditional algorithms,this method not only improves matching accuracy but also exhibits stronger robustness in complex environments.
作者
李忠良
黄河源
罗力
汤进东
陈星
LI Zhongliang;HUANG Heyuan;LUO Li;TANG Jindong;CHEN Xing(Fujian Polytechnic of Water Conservancy and Electric Power,Sanming,Fujian 366000,China)
出处
《自动化应用》
2025年第16期1-4,共4页
Automation Application
基金
2024年三明市引导性科技项目计划“基于深度学习的输电线路绝缘子缺陷实时检测算法改进研究”(2024-S-053)。
