Adaptive-scale determining for edge detecion in correlated texture noise

The selection of an appropriate filter size or scale is an important problem for edge detection. Using a single fixed scale over the entire image cannot produce desirable results: small-scaled filters are sensitive to step edges but have little control over noise, whereas large-scale filters have superior noise suppression characteristics, but they suffer from wide response around edges. This problem becomes more severe when the input images are corrupted by spatial correlated texture noise. In order to make the detection filter robust to texture noise, we tend to collect color information from neighbors as much as possible in homogeneous color areas and to reduce the number of neighbors when the locations close to edges. We introduce an adaptive scheme in this paper to determine an appropriate filter size or scale for each pixel base on local color statistics and hypothesis test on confidence intervals. This adaptive-scale determining scheme is integrated into a Canny edge detector, and the experimental results show the ability of the new procedure to detect both salient and weak edges from color textile images.