A Learning-Based Approach for Surface Defect Detection Using Small Image Datasets

Abstract Quality management is a fundamental component of a manufacturing process. In this paper, we propose a promising learning-based approach for automatic defect detection based on small image datasets. With the help of Wasserstein generative adversarial nets (WGANs), feature-extraction-based transfer learning techniques, and multi-model ensemble framework, our approach is able to deal with imbalanced and severely rare images with defects successfully, which is practically useful to the manufacturing industry. In addition, we reduce the false negative rate (FNR) as much as possible. Extensive experiments of defect detection on decorative sheets and welding joints achieve FNR accuracy results as 0.47% and 1.9% respectively, while traditional vision methods using in the production line can only achieve FNR results at about 20% under the similar circumstance, thus substantiating the proposed approach is quite effective for surface defect detection.