Steel surface defects classification method based on self-paced learning

Abstract

Steel surface defects are important factors that determine the quality of steel products. Automatic classification of steel surface defects can improve steel utilization and product quality. However, due to the low quality and small amount of steel surface defect data, the conventional classifiers are difficult to obtain a good classification result. Self-paced learning (SPL) provides a learning paradigm from simplify to difficulty. We combine the idea of SPL with DenseNet and propose the SP-DenseNet classification framework. A special self-paced process is designed for the surface defects of steel. In addition, the relationship between SPL and sample quality is further discussed. Experiments show that SP-DenseNet can solve the problem of classification of steel surface defects well. And for data sets with low sample quality, SP-DenseNet improves the classification results even more.