Local Method for Curved Edges and Corners

A new method of detecting features in Gaussian smoothed images is described. Applied to the simple case of marking the Canny edge pixels the method gives an improved response at comers where traditional methods have problems. Moreover, the method permits marking of more sophisticated features at essentially no extra cost. Additional information available includes local curvature estimates for each edge, direct (local) identification of occluded edges and marking and characterising comers on edges. The performance of the method on real images is compared with the PlesseyHarris comer detector. We find that the quality and sensitivity of these comer detection methods are similar, with the new method giving information on the orientation and opening angle of the comers. Finally it is shown how ellipse detection in noisy images is made possible using this curvature information. Introduction The Canny edge detector [1] consists in practice of the following: (a) Gaussian convolutions to yield x and y gradients of the Gaussian smoothed image (GSI) (b) Checking adjacent pixels in the GSI for maxima of the gradient magnitude in the (discretized) gradient direction. This step is a crude substitution for computing further derivatives, which is where our new method makes an impovement. At least two significantly different implementations of this scheme are in use. In one (eg the one originating at ATVRU Sheffield) only pixels above and to the side are used to perform step (b) above. In the other (eg from RAL Didcot) the diagonal pixels are used as well. Because these are discrete methods they have problems when the direction of the edge is changing quickly. The first approach leaves gaps at comers while the second gives spurious short edges. These problems at comers are not serious in a typical application because non-local processing will generally be applied to break up edges into segments at comers. However if one is hoping to extract a more detailed local characterisation of the image behaviour than just edge position and orientation then we need a more sophisticated treatment of the higher derivatives. New Computational Scheme The new method uses a continuous rather than a discrete image (as introduced in [3], [4]). We take as our Gaussian smoothed image the analytic function , y ) = X ' i j < *•' y-i) BMVC 1991 doi:10.5244/C.5.40