Long‐Tailed Steel Surface Defect Classification Using a Multi‐Expert Ensemble Framework

Source abstract

ABSTRACT In industrial steel surface defect classification, defect images often exhibit a pronounced long‐tailed distribution, where tail classes characterized by scarce samples and subtle visual patterns are much harder to recognize than head classes with abundant data. This imbalance typically leads to high misclassification rates and prediction bias toward head classes. To address this long‐tailed multi‐label classification setting, we propose a heterogeneous multi‐expert ensemble framework that integrates classification and detection experts. The classification branch further adopts a background‐aware Two‐Stage strategy to reduce background‐defect confusion caused by abundant defect‐free images. The framework exploits the complementary strengths of different experts, selects confidence thresholds and expert weights on the validation set and uses inference‐efficient experts for ensemble construction to reduce computational cost. Experimental results show that the method improves rare Class 2 recognition on the long‐tailed Severstal dataset, reaching an F1‐score of 0.912, while maintaining competitive macro‐averaged performance. Results on the balanced NEU dataset further show that the ensemble framework remains competitive under a different steel‐defect data distribution.