Development of methods and algorithms of technical vision for detecting the defect longitudinal crack on sheet metal

Abstract

The paper presents a mathematical model of a fuzzy subset of a defect in a digital image and is described as a piecewise constant function. The analysis of the filtering of the flaw detection image is given to ensure the implementation of the quantization algorithm of detection with subsequent adaptive binarization of the obtained result. The developed method makes it possible to detect a sheet metal defect of the longitudinal crack type and calculate various geometric parameters of this defect. This approach allows not only to see the detection of a longitudinal crack, but also to minimize the errors of the second level of false positives on flaw detection images. The above result is compared with the annotation of the flaw detector and with the YOLOv3 neural network.