Enhanced Defect Classification for Steel Plates via Magneto-Optical Imaging with Deep Feature Fusion

Steel plate structural integrity is vital for infrastructure and industrial applications, but surface/subsurface defects severely reduce component reliability. Conventional NDT techniques have limitations: Ultrasonic testing is sensitive to surface roughness, eddy current testing lacks deep flaw sensitivity, and vision-based methods are affected by illumination. Although MOI is promising for defect inspection, it faces noisy/low-contrast images and poor feature extraction; existing deep learning models struggle to balance accuracy and real-time performance. Herein, a real-time MOI defect detection framework with deep feature fusion under alternating magnetic excitation is proposed. It uses Pix2Pix cGAN for data augmentation, integrates S-GhostConv for efficient feature extraction, and adopts an improved PANet with attention mechanisms for multiscale fusion. Experiments on real and 6000 synthetic MOI images (four defect types) show it achieves 0.990 mAP0.5 and 130 FPS, outperforming YOLOv8s by 7.1% in accuracy. This framework provides a reliable solution for industrial steel plate defect inspection with broad application prospects.