Using gradient orientation for edgel detection

Abstract

Selection of suitable thresholds for edgel detection is discussed. An analysis of the variation in gradient direction is performed to estimate the probability density function of gradient direction as a function of image signal-to-noise ratio. A direction band about the mean of the PDF is determined on the basis of the minimum acceptable fraction of the cumulative neighborhood gradient magnitudes due to those gradients which have directions within the band. The proposed algorithm thus bases edgel detection on the distribution of local gradient direction rather than merely the magnitude response. Corners and small-radius curves are accommodated by permitting multiple histogram peaks with attendant reduction in both the minimum cumulative response fraction required and the direction bandwidth.<>