MGFP-Net: enhancing strip steel surface defect detection with multiscale ghostconv and feature pyramid shared block

Abstract In strip steel production, surface defect detection plays a crucial role in quality control and production efficiency. However, existing detection methods often struggle with the trade-off between computational efficiency and detection accuracy, particularly when handling diverse defect morphologies such as crazing, inclusions, and scratches under complex industrial conditions. To address these challenges, we propose MGFP-Net, a novel steel surface defect detection network incorporating three key innovations tailored for industrial inspection scenarios. Firstly, we introduce a Multi-scale Feature Extraction Block (MFE-block) that integrates our proposed Multiscale GhostConv (MSGConv) with an Effective Squeeze-and-Excitation (eSE) mechanism. This block employs an innovative channel grouping strategy, enabling efficient capture of both fine-grained texture variations and large-scale defect patterns in steel surfaces. Secondly, the Feature Pyramid Shared Block (FPS-block) is developed to optimize receptive fields through parameter sharing and progressive dilated convolutions, effectively addressing the varying scales of steel surface defects from microscopic cracks to macroscopic rolled-in scales. Thirdly, the Ghost-Lightweight Block (GL-block), integrated within the model’s neck section, reduces computational redundancy in feature processing while preserving crucial defect characteristics through efficient feature generation mechanisms. Comprehensive experimental results on the NEU-DET dataset demonstrate that MGFP-Net achieves a mean Average Precision (mAP@0.5) of 82.6%, significantly outperforming current state-of-the-art methods. Notably, our model maintains high processing speeds of 240 FPS with only 18.6M parameters, meeting the real-time requirements of industrial production lines. The proposed architecture demonstrates particular strength in detecting challenging defect categories, achieving 78.6% AP for crazing defects and 92.4% for inclusion defects. Cross-dataset validation on GC10-DET further confirms the model’s generalization capability, achieving 70.7% mAP@0.5 and outperforming baseline methods by 3.7% while maintaining superior computational efficiency. Moreover, it effectively handles complex defect distributions and background variations in strip steel surfaces.