A multi-class classifier based on support vector hyper-spheres for steel plate surface defects

Abstract In order to improve classification accuracy and efficiency for steel plate surface defects, a novel multi-class classifier termed as support vector hyper-spheres with insensitivity to noise (INSVHs) is proposed in this paper. On one hand, the INSVHs classifier introduces pinball loss to reduce its sensitivity to noise around decision boundary. On the other hand, the INSVHs classifier reduces the adverse effect of label noise and enhances the beneficial effect of important samples by adding local within-class sample density weight. Moreover, the INSVHs classifier builds an independent hyper-sphere for each type of defect to improve classification efficiency. The testing experiments for steel plate surface defects show that our INSVHs classifier is insensitive to noise and improves classification accuracy and efficiency.