Experimental comparison of image thresholding methods for defect detection in the papermaking process

Automatic detection and assessment of dirt particles in pulp and paper plays a pivotal role in the papermaking industry. Traditional visual inspection by human operators is giving the way to machine vision, which provides many potential advantages in terms of speed, accuracy and repeatability. Such systems make use of image processing algorithms which aim at separating paper and pulp impurities from the background. The most common approach is based on image thresholding, which consists of determining a set of intensity values that split an image into one or more classes, each representing either the background (i.e.: an area with no defects) or an area with some types of contraries. In this paper we present a quantitative experimental evaluation of four image thresholding methods (i.e.: Otsu’s, Kapur’s, Kittler’s and Yen’s) for dirt analysis in paper. The results show that Kittler’s method is the most stable and reliable for this task.