Kalman predictor based edge detector for noisy images

Edge Detection is a primary but one of the most essential segmentation tasks of image processing. Though numerous techniques are available for edge detection, it is hard to find a generalized version adaptive to all situations. Edge detection challenge gets stiffer in case of noisy images, because most of the derivative based edge detectors are very sensitive to noise. In this paper, we have tried to attack the edge detection problem from a different perspective. Instead of finding gradient, we run a Kalman Predictor over the image from two opposite directions of horizontal and vertical dimensions. Error between estimated and actual pixel values provides cue for edge localization, which is further processed by dual threshold to get the true edges. Proposed edge detector performs quite satisfactorily in case of noisy images and can be used for text extraction from noisy document image or medical images corrupted by artifacts.