Edge detection using scale space knowledge

Edge detection has acquired enormous importance in computer vision research: gaussian filter has been widely used in the past to accomplish this task. Although the gaussian filter has nice scaling behaviour and is computationally elegant, it suffers from the disadvantage of delocalisation of edges when operated at higher scales. Multi-scale processing of an image is thus required. In this paper, the process of edge detection has been looked upon as a reasoning problem. In the past, knowledge handling and reasoning have been attributed to high level vision routines, but from the results presented here, it can be argued that reasoning does play an important role in the process of edge detection. The knowledge base required for this is formed out of the theory of scale space. Particular emphasis is laid on the behaviour of delocalised, missing and false or spurious edges in the scale space. The scale space is formed by operating the LoG filter of different sizes on the input image. The dissimilarity among the zero crossings is measured across the scales. It is useful in the removal of false edges. A compatibility measure relates the zero-crossing contour with the current scale parameter, /spl sigma/. The rules are coded in NEXPERT OBJECT 2.0. The results are compared with Canny´s multiple edge detection algorithm (J.F. Canny, 1986).<>