Visual Perception Based Strip Surface Defect Detection And Recognition

As an important part of iron and steel products,the quality of steel production,which has become an important raw material for aerospace,machinery manufacturing,automobile manufacturing,chemical and other industries,will directly affect the final performance of the product.Thus,as an important method for automatic strip surface quality assessment,the research on the machine vision based on-line detection of strip surface defects has important theoretical and practical significance.In view of the conventional defect detection method implemented on the practical application is of low throughput and detection accuracy,based on the aspects of texture anomaly detection and machine learning,this paper presents two defect detect method—LBP based and CNN based defect detect method.Meanwhile,to increase the accuracy of the defect classification,an improved Relief F feature selection method is proposed.The contents and results of this paper are as follows:According to the high false detection rate and false negative rate of traditional steel strip defect detection method and the multi parameters,a new method based on multi-scale LBP encoding is proposed.Firstly,build a Gauss difference Pyramid model under different scales for strip steel image,and the initial suspicious areas will be found after this step.Then,implement threshold method and LBP encoding to the suspicious area.Finally,fuse all the encoding images.Then the detection results of the connected domain(ROI merge)are then generated to generate the complete defect information.In order to eliminate the interference of background texture and appearance of pseudo defects,a new defect detection method based on singular value decomposition is proposed,which is called SVD-LBPH in this paper.First,SVD is used to decompose and reconstruct the image.Then,LBP encoding method is implemented and after this step,LBP histogram based statistical features are calculated,compared them with a threshold,the defects are detected.Real time detection stage requires higher real-time performance.After real-time detection,instant-time detection is needed to detect defect in the defect area detected in the real-time detection stage.Considering the high accuracy of the target detection and image classification of deep learning algorithm,a convolutional neural network based defect detect method is Implemented.Comparing the performance with other machine learning based defect detect methods,experiments show that convolutional neural network is of high potential in defect detect task.In order to identify the defect types,gray features,gray level co-occurrence matrix and frequency domain features are extracted.In order to improve the accuracy of classification and remove invalid features,feature reduction method is adopted to avoid over fitting.In this paper,we use an improvement of ReliefF to reduce feature dimension.Finally,the reduction feature is classified by SVM.