Surface Defect Detection of Magnetic Tiles Based on an Improved Lightweight GhostNet Network

The performance of permanent magnet motors in modern industry is highly dependent on the quality of magnetic tiles, the surface defects of which can significantly degrade the efficiency and reliability of the motors. Traditional inspection methods are plagued by insufficient precision and stability, whereas deep learning-based approaches, despite offering superior accuracy, entail substantial computational overheads. To address these challenges, this study proposes a lightweight and optimized model, namely S-GhostNet, for the defect detection of magnetic tiles. By integrating multi-scale Ghost features through the adoption of dilated convolutions with different rates, channel shuffling, depthwise separable convolutions, and a Feature Pyramid Network (FPN), the model enhances the diversity and fusion of features while minimizing redundancy. Experimental results indicate that S-GhostNet achieves an accuracy of 95.46\%, representing a 14.36\% improvement over the baseline GhostNet, and accompanied by a 29.76\% reduction in computational cost (from 4.20 G FLOPs to 2.95 G FLOPs). These findings validate the effectiveness and efficiency of the model in industrial applications.