Edge detection comparison study and discussion of a new methodology

A comparison study of well known edge detection methods (Sobel, Prewitt, Canny, LoG, Zerocross, Roberts) for binary images revealed that these methods have a tendency to distort images, especially under noisy conditions, with some methods exhibiting image distortion even under noiseless conditions. While Sobel, Prewitt, and Canny performed better overall, it was observed that they do not completely filter out noise (Gaussian, Poisson, Salt & Pepper, and Speckle). In addition, the scores dropped dramatically when edges within 0 pixel offset were considered. A new edge detection algorithm was developed that gave significance to pixel distribution. A threshold value was defined below which the pixels were considered to be insignificant. Another key point of the algorithm was to find the proximity of the pixels relative to the computed constants alpha and beta, which helped to filter out noises such as Gaussian noise. The edge detection algorithm is simple-if there are at least c₁ pixels that have a different color (or shade) than the target pixel, and if there are at least c₂ pixels that have the same color (or shade) of the target pixel, then this condition indicates the presence of an edge. The first check is performed to see whether there is a color (or shade) variation near the pixel location, and the second check is performed to see if there is pixel continuity. These combinations, together, effectively indicate the presence of an edge.