Modeling and recognition of steel-plate surface defects based on a new backward boosting algorithm

Abstract

The surface quality of steel plates that have been widely used in the manufacturing industry directly affects the final product performance. The existing inspection system of steel-plate surface defects has some drawbacks: (1) an unbalance problem in the steel-plate surface defect dataset, and (2) the number of classifiers used in the recognition process is insufficient for identifying stains, dirtiness, and other non-common defects. It is imperative to develop a new method to identify steel-plate surface defects. In this paper, the normalization technique and the synthetic minority over-sampling technique (SMOTE) are used to establish a steel-plate surface defect dataset with complete categories, balanced quantities, and normalized features. Based on the existing boosting algorithms in the literature, a new backward AdaBoost (AdaBoost.BK) algorithm is proposed for defect recognition. AdaBoost.BK selects the most suitable weak classifier by the filtering mechanism, thus increasing the number of weak classifiers that can be combined. Experiments show that the model not only improves the recognition accuracy of non-common defects, but also improves the accuracy of the whole classification.