Surface defect detection of magnetic tiles based on an improved lightweight GhostNet network

Abstract The performance of permanent magnet motors in modern industry depends heavily on magnetic tile quality, as surface defects can significantly reduce motor efficiency and reliability. In magnetic tile surface defect detection, traditional visual methods have long been used but have limitations: they rely on manual feature extraction and rule-based judgment, showing poor adaptability to complex defect types and susceptibility to environmental factors like lighting. Thus, they hardly meet modern production’s high-precision and real-time demands, which previous studies only summarized briefly without in-depth analysis. Deep learning-based approaches have better detection accuracy, yet existing methods often require excessive computational resources—leading to large models and high overheads, which hinder deployment on edge devices. However, previous studies only ‘proposed methods’ without analyzing these drawbacks, offering little reference value. To address this, this study proposes a lightweight optimized model, S-GhostNet, for magnetic tile defect detection. By integrating multi-scale Ghost features via dilated convolutions, channel shuffling, depthwise separable convolutions, and FPN, it enhances feature diversity and fusion while reducing redundancy. Experimental results show S-GhostNet achieves 95.46% accuracy (14.36% higher than baseline GhostNet) and cuts computational cost by 29.76% (from 4.20 G FLOPs to 2.95 G FLOPs), verifying its effectiveness and efficiency in industrial applications.