Research On Strip Surface Defect Classification Based On Bag Of Visual Words

The technology of defect recognition based on vision has become one of the important technology of strip surface quality control.The advantages and disadvantages of the feature extraction method directly affect the recognition results.In this thesis,the characteristics of strip surface defects are analyzed.The feature extraction and classification based on Bag-of-Visual-Words(BoVW)model are employed.The existing BoVW model in the recognition of strip surface defects has problems,including uneven distribution of data,the ability to distinguish visual words,the information lack the overall structure of the target,and false defects interference.This thesis makes a deep analysis and puts forward the corresponding solutions.The specific research contents are as follows:(1)In order to test the performance of feature extraction methods,the strip surface defect image data sets HG-steel and PSC-steel were established.HG-steel set come from the production line of Handan Steel Corp.PSC-steel set come from the German PARSYTEL standard library.Among them,HG-steel set includes 2512 images,with 10 kinds of defects,such as drops tar,inclusions,pitted surface,and so on.PSC-steel set includes 2838 images,with 16 kinds of defects,such as hole,edge crack,waterarea,and so on.HG-steel and PSC-steel sets in the defect categories are not the same.So it can effectively achieve the complementary between the data sets,and fully test the performance of the feature extraction method.(2)In view of the problem of imbalanced data distribution,the visual codebook generation method under the guidance of classes is proposed.Since considering the information of defect categories in the visual codebook generation phase,the ability and the representation of the visual codebook are enhanced.And the problem of the lower classes recognition rate resulted from small number of data sets is effectively solved.In this thesis,the recognition rate of 96.95% and 97.72% on the two data sets are obtained.Experimental results demonstrated that the proposed algorithm can enhance the recognition rate of small sample defect categories,and effectively reduce the negative impact of data distribution imbalance on the performance of defect classification.(3)In view of the problem that the traditional BoVW model does not consider the ability of the visual word to represent the defect,a new feature representation method of weighted coding is proposed.The method adjust the weight of visual words in the visual codebook by calculating the frequency and information gain of visual words at the stage of feature representation.In this thesis,the weighted coding algorithm is used to increase the weight of visual words with high discrimination,reduce weight of visual words with low discrimination.The feature extraction method is more effective to distinguish the differences between the various types of defects.In this thesis,the algorithm is compared with the traditional BoVW model with equal weights in HG-steel and PSC-steel sets.The experimental results show that the proposed method better represent all kinds of defects.(4)In view of the problem that traditional BoVW model lacks the overall outline and structure information of the image,a new method based on local statistical features and global structure is proposed.In the traditional BoVW model,the global Gist feature is introduced to realize the complementation with local features.Experiments explained that a better result is obtained,and the validity of the method is verified.(5)Aiming at the problem of false defects in the strip surface background image,some false defects are simulated by Gauss noise.The robustness of the three kinds of classification algorithms to the noise interference is evaluated.The experimental results show that the proposed method can get better classification performance even if there are false defects interference.