Image segmentation by gradient statistics
Kenong Wu, Steven Schreiner, Brent Mittelstadt, Leland Witherspoon · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1998
This paper introduces a new gradient-based thresholding method for segmenting gray level images. This method first computes the magnitudes of image gradients. It, then, determines a range of threshold candidates from a statistic measure, called average of averaged gradients. Finally, it derives the image threshold from those candidates. The algorithm is fully automatic and does not analyze the shape of the image histogram. Unlike most gradient-based thresholding methods, this approach effectively reduces the influence of noise in both object and background regions to the threshold selection by computing the threshold from an intensity range, which corresponds only to the intensities at the boundary regions between the object and its background. It is more accurate and orders of magnitude faster than a similar approach. The experiments with synthetic images and real medical images are performed. Comparisons between this method and three other gradient- based approaches are conducted.