Research On Multi-class Classification Method Of Steel Surface Defects Based On Hyper-spheres

In recent years,China’s steel enterprises have been greatly improved in production equipment,production processes,types and quality of products.And the annual output often ranks the first in the world.However,steel products can’t meet the requirements of high quality from the market.The quality of steel plate surface is one of the important criteria to determine the rank of the steel plate.And the surface defect is the most important problem to hinder the steel enterprises to improve the quality of steel products.Therefore,the research on the detection technology of steel plate surface defect has become the key point.Most enterprises usually adopt surface defect detection system with visual acquisition device to detect steel plate surface defect.In the process of detection and recognition,defect classification is the central task.And it is one of the final goals of surface quality detection.In this paper,the classification model of steel plate surface defects is studied,and the main contents are summarized as follows:(1)In order to improve the classification accuracy for the steel plate surface defects,the classification model with hypersphere is preferentially used in this paper.Moreover,in order to restrain the influence of feature noise on the classification accuracy,a novel support vector hyper-spheres multi-class classification model(ASVHs)is proposed on the basis of the twin-hypersphere support vector machine(THSVM).In order to restrain the bad effect of boundary feature noise,the novel attribute is introduced into the classification model.The attribute changes the strategy that the classification hypersphere is built based on the boundary samples.(2)In order to solve the problem of large storage space and low efficiency caused by the large amount of defect sample data,a sample sparse strategy names as parameter iteration adjustment(PIA)is proposed based on the ASVHs classification model.The strategy uses the support vector data description(SVDD)to construct a hypersphere for each type of defect data.The size of the hypersphere is adjusted by controlling the penalty parameters in SVDD.Then the samples in the hypersphere are pruned.And the samples out of the hypersphere are retained.Moreover,based on the PIA strategy,the ASVHs classification model is rebuilt.Experiments show that the ASVHs classification model based on PIA reduces the cost of storage space and on-line operation time.(3)In order to further improve the operational efficiency of ASVHs,ASVH-WOA classification model is proposed in this paper.This model uses the novel whale optimization algorithm(WOA)to search the optimal three parameters for ASVHs.Experiments show that the novel method greatly improves the operation efficiency under the premise of high classification accuracy.