A model-based approach for filtering and edge detection in noisy images

The authors consider the problem of enhancement and edge detection on noisy, real-world images. The restoration and edge detection framework is based on an autoregressive (AR) random-field model. An edge is detected if the first and second directional derivatives and a local estimate of the variance at each point satisfy certain criteria. When noise is present, a good estimate of the original from the noisy images improves the signal-to-noise ratio, resulting in better estimates of the directional derivatives. To avoid excessive computation, the problem of estimation of the original image and the model parameters is presented as a combination of a reduced-update Kalman filter and an adaptive-least-squares parameter estimation algorithm. The restoration process is completed with a min-max replacement scheme to enhance edge strength. An orientation-sensitive detector resulting from the use of an AR model may not detect edges of significantly different orientations. This is partially overcome by running four edge detectors on the four interior pixels of a 4×4 window; this corresponds to rotating the window in successive multiples of 90°. Comparisons with R.M. Haralick´s (1984) facet model edge detector, R. Nevatia and K.R. Babu´s (1980) line finder, and J. Canny´s (1986) edge detector are given