Research on Target Detection Algorithm Based on YOLOv8n-MtCW for Metallic and Non-metallic Impurities in Steel

In the process of steel manufacturing, impurity detection plays a critical role in ensuring product quality. However, due to the complex morphology of impurities, the diversity of impurity types, and the challenges posed by small targets and varying backgrounds, accurate detection remains a significant challenge. To address these issues, this paper proposes a novel target detection solution for metallic and non-metallic impurities in steel, based on the YOLOv8n-MtCW model, which integrates key optimization techniques for feature extraction and image processing in complex scenes. First, we designed a multi-task concurrent detection network that leverages label decoupling and specialized migration learning to enhance the model’s adaptability to diverse impurity types. This network employs different detection sub-networks for impurities in distinct backgrounds, improving detection capabilities across various scenarios. Additionally, a Convolutional Hierarchical Attention Module (CHAM) is introduced to enhance the model’s ability to extract global and local features, thereby improving its overall feature representation. Finally, we implemented the Wise-IoU loss function to suppress the impact of low-quality samples during model training, improving the detection accuracy for small and complex impurities. Experimental results demonstrate that the proposed method achieves a mean Average Precision (mAP) of 62.3%, outperforming the standard YOLOv8n model by 10.4%. These findings validate the effectiveness of the proposed method and provide technical support for impurity detection in steel manufacturing.