Statistical discriminator of surface defects on hot rolled steel

A statistical approach to defect detection and discrimination has been applied to the case of hot rolled steel. The probability distribution of pixel intensities has been estimated from a small set of images without defects, and this distribution is used to select pixels with unlikely values as candidates for defects. Discrimination of true defects from random noise pixels is achieved by a dynamical thresholding procedure, which tracks the behaviour of clusters of selected pixels for varying threshold level. Boundary levels of the dynamic threshold range are determined from the estimated probability distribution of the pixel intensities.