Skip to Main Content
Edge detection is one of the most commonly used operations in image analysis. Most algorithms contain two basic steps: denoise and derivative computing. We apply kernel regression to remove noise and to get gray-level and derivative intensity surface of images. We explore the Nadaraya-Watson kernel regression which conquers the more negative impact caused by noises for derivative computing than general algorithms. However it also smoothes the jump points which may be edge pixels in image. So we also present bilateral kernel regression to deal with the problem. Experiments are carried out for extracting edge information from real images, without and with the contamination of Gaussian white noise. For each sample image, edge is extracted under the two cases without noises and with the peak-signal-noise-ratio (PSNR) from 18.5 dB to 34 dB. The proposed algorithm is compared with several other existing methods, the Prewitt and Canny detectors. The experimental results indicate that our method has a better performance in noisy images than other methods.