1 Layout

A new zero-crossing edge detection method is described. Similar to Laplace of a Gaussian (LoG) algorithm, a 7x7 point spread function (PSF) is convolved with the original image to generate a smoothed image. Instead of locating edge position from the smoothed image, as in the case of LoG, we use the smoothed image as a reference to convert the original image into a three-value edge polarity map (negative edges, positive edges, and zerocrossing pixels). We show that an optimal threshold can be derived from the analytical study of the gray value difference between the original and the reference images, taking into account four edge models (ramp, step, roof, and gate). This threshold is determined such that the zero-crossings in the polarity map are of single pixel width. Finally a corner preserving detector with a moving 5x5 window is implemented to identify real zero-crossings. The advantage of the proposed method over LoG is that it is more robust in preserving edges around sharp corners and thin lines. The new method is also 40% faster than the LoG. However, the LoG performs better for noisy, low contrast images. Experimental results with a test pattern image, a normal and a low contrast images are presented.