文摘
Automatic detection of defects on the surface of products or raw material is an important task in the field of automated visual inspection. Thresholding is a method for image segmentation that is often used for the detection of said defects. Several methods to select the optimal thresholding values automatically on a per-image base have been described in the literature. Some of these are particularly designed to deal with mostly homogeneous images such as those of product surfaces with some defects, but have not been tested sufficiently and not in the context of automated visual inspection. In this work we present a comparison based on such experimental conditions by means of an automated visual inspection station and a set of images specially acquired for this purpose. The methods that were compared are: the Otsu’s method, the Valley-Emphasis method, the Valley Emphasis with Neighborhood method, the Kittler-Illingworth’s method, and the Maximum Similarity Thresholding method. The highest performance, with a statistically significant difference, was obtained by Maximum Similarity Thresholding.