Steel strips surface classification and defect detection based on multiple integrated features and classification schemes

Machine vision has become an indispensable tool in automated steel surface inspection. Such technologies are able to facilitate or replace manual inspection methods with benefits such as manpower reduction and operator error minimization. In the last two decades, researchers have actively explored computer extractable visual features for steel surface inspection. However, existing approaches suffer from certain limitations that cause ineffectiveness. Specifically, (i) non-discriminating feature choices lead to poor inter-class separability; and (ii) classifier complexity coupled with high numbers of training epochs. Therefore, this research aims to propose two frameworks to improve the inspection performance of steel surface types. The first framework performs two tasks for machined surface texture classification and identification. The first task generates the most discriminating feature representation for surface texture. This is achieved through the proposed feature extraction method DST-GLCM, which integrates the Discrete Shearlet Transform (DST) and the Gray Level Co-occurrence Matrix (GLCM), producing a compact yet discriminative feature representation. A two-level classification scheme is then proposed combining the capabilities of the Support Vector Machine (SVM) and a proposed Consecutive Training with Collective Testing Artificial Neural Network (CTCT-ANN) technique. The SVM-level classifies the surface images into six categories based on surface texture features. This is followed by the CTCT-ANN-level where each of the surface textured images are further classified into sub-categories according to their surface roughness. Finally, the surface roughness parameters for all classified images are estimated. The second framework extracts and combines different features (global and local) from hot-rolled steel surface images of different forms into different domains (spatial and frequency). This improves the image content description and offers a variant representation of the surface image. This is useful for surface defect detection and classification. The global features are extracted from the input image using the proposed DST-GLCM (from Framework-1). Local features are extracted after dividing each input image into four blocks. Then, local features/descriptors namely the GLCM, Uniform Local Binary Pattern (ULBP) and Speeded-Up Robust Features (SURF) are extracted from every block. All the extracted global and local features are combined in a high dimensional feature vector, whose dimensionality is later reduced using Principal Components Analysis (PCA). The final classification is accomplished using an SVM. Both frameworks are evaluated using two different datasets of steel surface images. The first framework uses Engineering Machined Textures (EMT) workpiece surface images produced using several machining processes. The second framework uses the Northeastern University (NEU) standard database that comprises surface images of hot-rolled steel strips with different defect types. The results in this research show improvement when compared with previous related studies. The maximum accuracy achieved in surface classification of EMT dataset was up to 100%, with a maximum error in surface roughness estimation of 0.004 micrometer. In addition, the maximum accuracy achieved in defect detection of NEU dataset was up to 99.34%.

Text FSKTM 2018 80 - ir.pdf Download (971kB)

Actions (login required)

View Item