Estimating Angles And Curvature Features In Grey Scale Images

This paper presents a quantitative analysis of the error of the edge angle measurement obtained from edge detector in the presence of grey level gaussian noise. Our analysis shows that the uncertainty of the angle measurement is only related to the grey level noise power and the gradient of the edge point at which the angle is measured. Both image independent upper bound and image dependent estimate of the noise variance of the angle measurement are given. With this result, we argue that one can provide better angle and curvature estimates by filtering the measurements adaptively according to the local uncertainty without having to compromise detail sensitivity and noise robustness. As an example of application, a new method of detecting curvature features is derived. The experimental results show that this algorithm works well, and in particular it can handle some difficult situations where other methods may fail.