基于多尺度离散曲率的图像角点检测方法  

Image corner detection method based on multi-scale discrete curvature

作　　者：景军锋[1] 孙久锐 章为川[1] 白萌萌 JING Junfeng;SUN Jiurui;ZHANG Weichuan;BAI Mengmeng(School of Electronics and Information, Xi’an Polytechnic University, Xi’an 710048, China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《西安工程大学学报》2021年第5期27-33,共7页Journal of Xi’an Polytechnic University

基　　金：国家自然科学基金面上项目(62176204);陕西省教育厅服务地方专项计划项目(19JC018)。

摘　　要：传统的曲率计算方法对离散域上的局部变化和噪声敏感,使得角点检测准确率和定位存在偏差。为了减少上述问题,提出了一种多尺度离散曲率来计算曲率并检测角点。首先,用Canny边缘检测器处理输入图像,得到一个二值化边缘图,然后从边缘图中提取边缘轮廓,沿着轮廓填充缝隙,检测并标记T型角点;其次,在3个不同的平滑尺度下,用高斯函数对每条平面曲线进行平滑处理,再使用新的曲率度量方式,计算经平滑后每个轮廓上点的曲率,并将绝对曲率的局部极大值点作为各尺度的候选角点;最后,将候选角点曲率值与曲率阈值比较,大于曲率阈值的点定义为候选真实角点,并将3个尺度下均可检测到的候选角点作为真实角点。实验结果表明:本文算法对标准角点数据实验室图进行检测时,检测到的正确角点数为159,错误角点数为90,丢失角点数为16,定位误差为1.2572;对标准角点数据积木图进行检测时,检测到的正确角点数为46,错误角点数为13,丢失角点数为1,定位误差为1.0132。与现有的3种经典角点检测算法相比,本文的检测方法具有更好的角点检测性能。The traditional curvature calculation method is sensitive to local changes and noise in the discrete domain,which makes the corner detection accuracy and location deviate.In order to solve the above problems,a multi-scale discrete curvature is proposed to calculate the curvature and detect corners.Firstly,Canny edge detector was used to process the input image to obtain a binary edge image,then the edge contour was extracted from the edge image,the gap was filled along the contour,and T-shaped corners were detected and marked.Secondly,under three different smoothing scales,Gaussian function was used to smooth each plane curve,and then a new curvature measurement method was used to calculate the curvature of each point on each contour after smoothing,and the local maximum point of absolute curvature was taken as the candidate corner point of each scale.Finally,the curvature value of candidate corners was compared with the curvature threshold.The points greater than the curvature threshold were defined as candidate real corners,and the candidate corners that could be detected at three scales were regarded as real corners.The experimental results show that when the algorithm is used to detect the standard corner data laboratory image,the correct corner number is 159,the wrong corner number is 90,the missing corner number is 16,and the localization error is 1.2572,and that when the standard corner data block image is tested,the detected correct corner number is 46,the wrong corner number is 13,the missing corner number is 1,and the localization error is 1.0132.Compared with the existing three classical corner detection algorithms,the detection method in this paper has better corner detection performance.

关 键 词：角点检测 多尺度离散曲率 多尺度曲率尺度空间 特征提取 边缘检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]