Inspection And Identification For Band Steel Surface Defects

With the rapid development of productivity, people’s living standard has been greatlyimproved continuously, people make higher requirement on manufacturing industrycorrespondingly. As the essential raw material for shipbuilding, space flight&aviation and dailylife, the standard of steel quality is certainly high. The steel products quality and productioncapability are far from satisfying the demand due to the limitation of existing productionequipment and technology. On the band steel production line, the lack of effectively supervisionon the surface quality in the process band steel products production make some flaw productscannot be picked out before leaving factory, which will lead to huge economic loss to thecompany once the flaw products flow into the market. In addition, there is a chance for the newtypes of flaw in the practical application of band steel production, which bring the impact ondiscriminating accuracy of the traditional classification model.This thesis proposes the theory of real-time monitoring the production line on the basis ofmachine vision. The steps of the image processing are as the following: catching images viahigh-speed industrial camera firstly, fast detection and preprocessing on the images, segmentingthe defect area, and then using defect detection technologies to binary operation on the targetimage, locating the defect position, and getting the relevant attributions on the target and thebinary images, using the training model by making use of the machine learning methods to aclassification model, this can not only improve the productivity, but also reduce the laborintensity. And also the one-class algorithm based on the minimum spanning tree is aclassification to refuse one kind category depending on the covering model, which can be used tojudge a new category; combine the minimum spanning tree and SVM to get the newclassification algorithm to separate the new category efficiently to improve the classificationaccuracy. The results show that the proposed algorithm has better classification performance.