A Multi-Expert Ensemble Model for Long-Tailed Steel Surface Defect Detection

Yang DanDan, Ling Zhou

Keywords: Steel Surface Defect Classification, Long-Tailed Distribution, Multi-Expert Ensemble Model

Abstract: In the field of industrial steel surface defect detection, defect images often exhibit a pronounced long-tailed distribution, where tail-class—characterized by scarce samples and subtle features—are much harder to recognize than head classes with abundant data. This imbalance typically results in high miss-detection rates and bias toward head classes. To address this challenge, we propose a Multi-Expert ensemble model that integrates classification and detection experts, introducing a Two-Stage strategy into the classification branch. The framework leverages the complementary strengths of various experts, with validation-based joint optimization of confidence thresholds and expert weights, and employs parallelized training and inference to improve computational efficiency. Experimental results show that the method significantly improves the F1-score(0.912) of tail-classes 2 and achieves state-of-the-art average Accuracy(0.989) on the long-tailed Severstal dataset, while strong performance on the balanced NEU dataset further validates its cross-distribution generalizability and practical applicability for industrial steel surface defect detection.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 8467