Study On Surface Defect Detection System For Steel Strip Based On Image Processing

Quality control is an important guarantee for a enterprise to strengthen its competitive-ness, establish enterprise brand and inherit enterprise culture, especially in the iron and steelindustry. Living in "the winter of the steel industry", the only way is to strive for theenterprise transition, research and develop high-tech products and transfer from extensiveproduct to high-grade, precision and advanced product, in the face of various dilemma such asthe falling of profits, overcapacity and vicious competition in the industry. Traditional manualdetection methods is not only inefficient and having high rate of false alarms, but alsoconsuming large amounts of labor. Traditional nondestructive testing technology, such aseddy current testing technology, infrared detection technology and magnetic flux leakagedetection technology, can only detect limited defects types and the quantitative descriptionparameters of defects, and can not assess the surface quality status comprehensively, due tothe limitations of the detection principle. In recent years, with the development of computertechnology and automation technology, as well as the maturity of artificial intelligence andneural network theory, techniques of surface defects detection for steel strips, which have thecore of machine vision technology, have become the focus of current research.This dissertation does researches on some key technologies of the hot-rolled steel surfacedefects detection. Through this dissertation, the following achievements have been made:（1）According to the site environment and the needs and budget of the enterprise, thehardware and software design proposal for the hot-rolled steel surface defects detectionsystem have been decided.（2）According to the state of the environment of the hot-rolling workshop and theeconomic budget, a reasonably light source and lighting solutions have been chosen.（3）Proposed a composite filter, which was comprised of the mean filter and weightedmedian filter, on the basis of detailed analysis of the various filtering algorithm.（4）Determined a group of feature vector, composed of gray characteristics, texturecharacteristics and invariant moment characteristics, as the input of the defect classifier, afterthe study and analysis of the various types of characteristics of the steel plate surface defectimages.（5）Proposed an improved BP neural network algorithm, after the detailed study of thefour main pattern recognition methods and the selective analysis of the neural network patternrecognition methods.