一种大噪声自适应的角点检测技术  

A corner detection method based on strong noise adaptation

作　　者：杨栋[1] 尹义龙[1] 朱明英[2] 曹庆峰[3] 

机构地区：[1]山东大学计算机科学与技术学院,济南250101 [2]烟台南山学院软件工程学院,龙口265713 [3]山东大学工程训练中心,济南250061

出　　处：《南京大学学报（自然科学版）》2008年第2期140-147,共8页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(49672102)

摘　　要：在自主机器人视觉中,大噪声条件下图像边缘角点的识别,对于视觉目标的形状识别有着重要的作用.使用最小均方误差准则,可以有效的滤除噪声的影响.针对大噪声条件下噪声强度的平均值和均匀性都未知的特点,提出了一种基于最小均方误差的新的大噪声自适应角点识别算法.这种方法计算变长点集的残差平方和,使得角点出现在特征曲线的极小值处,避免了使用阈值的弊端.并利用折半查找策略,降低了时间复杂度.在单角点识别的基础上,针对多角点识别,讨论了利用这种算法降低时间复杂度的方法.通过分析表明这些算法能够较为有效地提高大噪声条件下单调边缘单角点识别的正确率,降低了时间复杂度.Corner is the point which has the sharp change of brightness in image, or the maximum curvature on contour. Accordingly, there are two kinds of corner detectors, contour-based ones and image-based ones. Contourbased corner detection plays an important part in shape recognition. Generally, there are three kinds of contour- based corner detectors, one based on maximum curvature, one based on beeline or curve fitting and the other one based on multiple scale spaces. However, up to now most of them don’t work well under strong noise conditions. Under strong noise conditions, we suppose the average and uniformity of noise are unknown. It is mainly because of complex vision conditions, such as camera noise, complex illumination and mechanical vibration of camera. To detect corner adaptively when noise conditions are unknown, we can simulate human vision system. According to Gestalt vision theory, human vision is of good continuity, and can filtrate continuous noise details adaptively and gain glancing corners of contour. When human vision searches corners, it can neither compute the curvature of everypoint, nor find the optimal criterion globally. Someone points out that human vision is sensitive to beeline section. And the least square method （LSM） is widely used for beeline fitting. However, most of beeline-fitting-based corner detectors cannot work well when noise level changes. The main reason is that they assume the noise level is small or fixed and known, so they use a user-defined or fixed threshold to detect corners. This paper points out that a corner is the point or points-set, the forgoing and following points set of which are most fitted with two beeline sections. And we use the whole points of local contour to detect a corner. Therefore, if the detail noise increases, it can only reduce its contribution to corner eigenvalue. There is the shortest distance between corners （SDBC）, under which the two corners can be recognized because of the strong noise. In the head and tail of a contour there are

关 键 词：机器视觉 大噪声 角点识别 最小均方误差 折半查找 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]