Research On Surface Defects Detection Theory And Recognition Algorithm Of Steel Strip By Images

Rolled steel strips are important materials in industrial productions, so the quality of steelstrips will have serious influence on that of end products.Surface defects on steel strips aresignificant factors in evaluating quality of them. Except improving rolling technology todecrease defects,Detecting defects in time to avoide application of defects’ strip in endproducts is also very important.Algorithms and theories of defects detecting on rolled steelstrip surfaces by images are researched in this article.The research topics are as follows：Quick defects detecting method is brought forward based on the principal that defectsdetecting and classification carried out in stages.Image complexity is introduced to quicklydetect defects on steel strip surfaces.Proceeding from reliability and time saving, four imagecomplexity describling parameters are compared to each other in detecting results, finallyRPVC is selected to detect defects on strip surfaces. Simulation shows that RPVC can detectdefects correctly using shorter time and realize detecting in time.Defects images preprocessing is also researched in this article, in view of the influenceon defect segmentation by textures on images’ background, combination of local complexityand local variance is applied to weaken background textures and Gaussian high frequencyfilter is used to homomorphicly filter defects’ images to wipe of bright spot and enhancecontrast. Three parameters are selected to evaluate the quality of filtered images andcompared to the results of histogram enhancing, all shows better effects.Visual features selection method is put forward to segment defect images by visualattention mechanism. After analysis of defect images, brightness, rareness and localcomplexity are selected to be features to generate salient images, expressions of them areprovided. Gaussian filter is used to get global salient images.Improved salient images fused method is brought forward. We modified the weightcoefficients of feature salient images fusion computering method, Normalized complexity andentropy values of salient images are used as weight to generate comprehensive salient image,this can reflect the differences of different salient images’ contribution to comprehensivesalient image. Maximum entropy is used to segment comprehensive salient images,compare with results of K-means method and region-growing method, the improvedsegmentation method based on visual attention model has better effect.For the purpose of defects classification, defect images’ features selecting method isprovided after analyzing the adaptability of different characters of them. In view of differentinformations included in unsegmented and segmented images, texture features, invariant moment features and dispersion are selected to be classification features.Extenics theory is introduced into images classification.Feasibility of extenics theory inimages classification is discussed. Weight coefficients of defects images unspecifided topre-selected types of defects images in computering relevancy values are improved and theexpression is provided, the effectiveness of the improved computering method is alsodemonstrated. The quotients of distance between feature values and classical domain andsum of these distances are used as weight coffecients in computering comprehensiverelevancy values.This computering method enhances the influences of defects’ self featurevalues on comprehensive relevancy..Classical domain and joint domain are key parameters in application of extenics theory,they are gotten by statistics analysis of a number of defects images.Seven defects types areselected to simulate the method, maximum relevancy value is used to group the unspecifieddefect image in one of preselected defect types, classification accuracy is close to90%,comparing to primary weight coefficient computering method, the improved theory is moreeffective in defects images classification.