A Non-Invasive Technique for Online Defect Detection on Steel Strip Surfaces

During the production of steel strips, a large amount of surface defects can be generated, due to harsh environmental conditions. A high number of surface defects can lead to rejection by the customer, which represents significant economic losses to the production plant. Thus, it is very important to detect the presence and type of the defects generated during the production of each steel strip. Using this information, it is possible to determine whether a strip is suitable for sale, and it may also be useful to determine the origin of defects and, if possible, prevent them from being generated in subsequent strips. To perform these tasks, non-invasive inspection techniques are usually used, carried out automatically by artificial vision systems. Although the inspection conducted by humans is more accurate, they become fatigued quickly, or may even be unable to carry out the inspection correctly when the forward speed of the strip is high. In this paper, a new detection technique is proposed, based on the division of an image into a set of overlapping areas. The optimum values for the configuration parameters of the detection technique are automatically determined using a genetic algorithm. After the detection phase, all the defects are classified using a neural network. A very satisfactory success rate has been achieved in both detection and classification phases.