Fast Defect Detection Algorithm For Surface Of Steel Strip From Coarseness To Fine

The quality of strip products has become more and more important to the final performance of other products.This paper presents a framework for the detection of defects from coarse to fine.It can effectively solve the problems of slow detection speed and high false detection rate in the existing defect detection methods,and can effectively remove noise samples and redundant features.The concrete contents and results of this paper are as follows:(1)In order to suppress the generation of false defects in the background image,the method based on anisotropic diffusion is used to preprocess the image.This method can weaken the background texture of the image while keeping the edge information of the defect.For the defect target with large gradient at the edge,the diffusion coefficient of the model is small,and the diffusion coefficient is relatively large in the background of the image with a small gradient.Experimental results show that the method can effectively suppress the interference of background textures and the generation of false defects.(2)To improve the detection speed,a method based on E-BING algorithm is proposed for predicting the surface defects of steel strip.Aimed at the problems existing in the strip surface defect detection based on the traditional BING algorithm,this paper integrated the entropy feature into the BING framework,and the candidate targets are selected under the guidance of the binarized normed gradients feature and the entropy feature.Experimental results show that the E-BING algorithm can effectively reduce the number of prediction windows while ensuring the quality of the candidate targets.(3)In order to obtain the compact features representation of the defects and solve the problem of noise interference in the training samples,this paper proposes a method of defect feature selection based on R-AdaBoost,and the method can effectively eliminate the noise samples in the training samples while the feature selection is carried out.The algorithm is first used to select features and construct effective and compact feature vector via Relief feature selection algorithm in each cyle of AdaBoost algorithm integration framework,and then remove noise samples according to intra class difference among samples in each cycle,and finally update training samples according to the sample weight.The algorithm is tested on Hangang defect image library,the test results show that compared with other algorithms,this algorithm improve the accuracy rate by about 2%,the number of features is reduced by 6.3%~9.4%,and verify the feasibility of the method.(4)In order to solve the contradiction between the detect detection efficiency and the detect detection rate,this paper proposes a method of BING-R-AdaBoost defect detection from coarse to fine.First,the E-BING fast prediction algorithm is used to predict the candidate target,and R-AdaBoost is used to analyse and extract the accurate feature from the target in the candidate window,and finally obtain the exact location of the defect region in the image.The test results on Hanganng defect image library show that the proposed algorithm not only reduce the error rate,but also can effectively improve the speed of detection.