基于ROC分析的Canny算法在景象匹配中的应用

ROC-based Canny algorithm for scene matching

根据ROC分析方法能对分类识别算法进行多门限评估的特点,利用其改进Canny算法性能并应用于景象匹配。首先由不同参数组合的Canny算子计算图像的多个边缘提取图,并逐像素进行统计,得到边缘像素相关图;然后采用ROC曲线分析找到最佳的关联阈值,从而确定理论边缘图;最后将参考图与实时图所对应的理论边缘图进行景象匹配。实验结果表明,该方法在参考图和实时图存在一定几何畸变和灰度差异的情况下,能取得较高的平均匹配精度与正确匹配概率。此外,该方法克服了传统Canny算子采用固定参数的缺点,根据多参数自动进行筛选优化,有较强的工程实用性。

Receiver Operating Characteristics （ROC） analysis method, which can perform a multiple threshold evaluation without the prior probability of classes, was applied to improve the performance of Canny edge detection algorithm in scene matching. Firstly, several edge detection results were computed using combinations of the Canny detector＇s parameters. The results were then tested for correspondence by each pixel. Then ground truth edge image was estimated using correspondence threshold which was optimized by ROC analysis. After obtaining the ground truth edges of reference image and real-time image, scene matching process was performed finally. The experimental results show that the method is robust and can be obtainable for high matching precision and registration probability at the condition of existing geometry distortion and luminance difference between matching images. This method has practical implementations of scene matching with automatic parameters setting process.