YOLO-MT: real-time and efficient surface defect detection method for magnetic sheet

Surface defect images of ferrite permanent magnetic sheets typically exhibit low contrast, complex textures, uneven illumination, and diverse defect types. To meet real-time detection requirements in industrial production, this paper proposes YOLO-MT, a novel defect detection framework tailored to the imaging characteristics and defect patterns of magnetic sheets. First, a C3k2_Her module integrating the proposed EdgeRefinement and Her (HierarchicalEdgeRefinement) structures is introduced to enhance backbone feature extraction and improve subtle defect recognition while reducing computational cost. Second, the Multi-layer Feature Pyramid (MFP), built using CSP_OmniKernel and SPDConv, strengthens global–local feature representation and boosts small-object detection. Additionally, a C3k2_FFCM module is incorporated into the neck to exploit frequency-domain global modeling for more accurate defect feature capture. A dedicated industrial magnetic sheet defect dataset is constructed, and extensive experiments validate the effectiveness of YOLO-MT. The proposed model achieves an mAP@50 of 97.2%, outperforming the baseline by 1.3% and exceeding the performance of mainstream detectors. YOLO-MT has been successfully deployed in real production environments, satisfying industrial requirements for accuracy and real-time detection.