In recent years, steel surface defect detection based on machine vision has attracted significant attention and has emerged as a research hotspot. However, several challenges remain. In practical industrial scenarios, deep learning-based detection methods often involve high computational complexity, which limits their applicability for real-time defect monitoring. Moreover, due to the complex and noisy background of steel surfaces, conventional deep learning networks frequently suffer from the loss of critical defect features during the feature extraction process. To address these challenges, this paper proposes a novel latent-space attention multi-scale YOLOv10n model (LAM-YOLOv10n). First, a lightweight ghost module is integrated to significantly reduce the model's parameter count and computational cost. Second, a spatial multi-scale attention (SMA) module is designed to enhance the extraction of discriminative features related to steel surface defects. Finally, a multi-branch feature fusion network (MFFN) is introduced to improve the effectiveness of multi-scale feature aggregation, thereby enhancing the model's detection performance for various defect types. Experimental results demonstrate that the proposed LAM-YOLOv10n model achieves a 3.47% improvement in precision compared with the baseline YOLOv10n network, outperforming several state-of-the-art object detection models in both accuracy and efficiency. These findings indicate the effectiveness and practicality of the proposed method for real-time steel surface defect detection in complex industrial environments.
Zhang et al. (2025) studied this question.