Multi-class classification for steel surface defects based on machine learning with quantile hyper-spheres

Abstract

Focusing on steel surface defects, a novel multi-class classification method is proposed. The method is termed as machine learning with quantile hyper-spheres (QH-ML). In order to obtain sparse set with boundary information from finite defect dataset, a new quantile hyper-sphere data description (QHDD) model is proposed. This model is used to generate a quantile hyper-sphere for each finite defect subset. And this quantile hyper-sphere is insensitive to noise. Then, in order to realize incremental learning for new samples, an incremental learning with quantile hyper-spheres (QHIL) method is proposed. The advantage of QHIL method is that the dataset is invariant in size during the process of incremental learning for new boundary information. In the meanwhile, a novel classifier with multiple quantile hyper-spheres (MQHC) is used to realize multi-class classification for steel surface defects. The target class of MQHC uses QHDD model, and negative class applies the margin maximization principle. MQHC has natural multi-class classification gene and perfect classification performance. In testing experiments, the proposed QH-ML is used to classify six types of defects with incremental learning. Experimental results show that QH-ML keeps high classification accuracy and efficiency.