EAAI Journal 2026 Journal Article
MCD-YOLO: An improved method for detecting mixed-class defects on steel plate surfaces
- Jiangwei Li
- Maoxiang Chu
- Peinan Zong
- Simin Ma
In the task of steel plate surface defect detection, existing detection models generally suffer from defect miss-detection issues caused by category prediction errors. To address this problem, this paper constructs mixed-category defect samples by fusing multi-category defects, and proposes a novel detection model — mixed-category defect YOLO (MCD-YOLO) — with such defects as the detection targets. The specific improvements are as follows. First, the reparameterized convolutional shuffle multi-scale dilated attention (RSA-MSDA) module is designed. This module enhances the feature extraction capability for mixed-category defects and simplifies the model inference process. Second, a vision Transformer for refined modeling of mixed-category defects is developed, namely the omni-dimensional dynamic convolution outlook (ODOutlook) attention module. Third, we integrate RSA-MSDA and ODOutlook attention into the C3K2 backbone structure to form RSA-MSDA-enhanced C3K2 (C3KF) and ODOutlook attention-enhanced C3K2 (C3KT) modules, boosting cross-level feature propagation for complex defect patterns. Then, a cross-level dynamic feature fusion mechanism named triplet attention selective feature fusion (TASFF) is proposed to improve the utilization rate of multi-scale defect features. Finally, a lightweight and efficient upsampler called dynamic upsampler (Dysample) is introduced. Experimental results demonstrate that, on three steel plate defect datasets, the proposed network reduces the average missed detection rate by 6. 0 percentage points and improves the mAP@50 by 7. 0 percentage points compared with the baseline YOLOv11, while maintaining the original detection efficiency.