Edge detection is a fundamental step in Computer Vision. Several edge detection schemes have been proposed in the computer vision literature. Some of the prominent ones include the Marr-Hildreth edge detector, the Haralick edge detector, and the Canny edge detector. Previous work by Ramesh and Haralick, ([9],[11]), and Wang and Binford [15] have provided theoretical and empirical evaluation of some of the edge detection schemes. This paper shows how we use the insights gained during the performance evaluation to develop a better edge detector. We illustrate that the precision of the edgel orientation estimate is a function of the input signal to noise ratio (the ratio of the true gradient magnitude to the gray level noise standard deviation) and the neighborhood size used in the edge detector. This observation has direct impact on edgel orientation estimation. We also illustrate that an appropriate measure for a pixel being an edge pixel along a given direction is the integrated gradient along that direction. We use the point that the minimum and maximum integrated gradient magnitudes are simultaneously high at edge locations to detect edge pixels. The paper also provides theoretical and empirical analysis of the performance of the operator.
展开▼
