A statistical approach of multiple resolution levels for canny edge detection

Vision processing needs effective feature detectors to estimate the structure and properties of objects in an image. The best known is Canny edge detection that combine a Gaussian low pass filter for noise reduction and non-maximal suppression and hysteresis threshold for edge localization. A possible problem of this approach is that the threshold values. Applying a single fixed threshold to gradient maxima is not an optimal choice. Thus, Canny uses two thresholds values namely T_low and T_high to reduce the number of false positive of pixels that represent significant contours in the image. However, by introducing two fixed threshold values are also not an optimal choice due to high variations in images. In this paper we introduce a method that computes the threshold values from the foreground and background image pixels. According to this method, an image is divided into several blocks using at multiple resolution levels. After that, a sampling approach is used on global and local regions to get the optimal thresholds by selecting the highest between class variance values. We have performed experiments on 200 images from the Berkeley dataset. The results show that the proposed method outperforms Canny that uses two fixed threshold values.