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

In modern industry, the performance of permanent magnet motors is crucial to enhancing efficiency and system reliability. As a key component of these motors, the quality of magnetic tiles directly impacts motor performance. Therefore, ensuring that magnetic tile surfaces are free of defects is essential for maintaining product quality. While traditional visual inspection methods have been widely used, they suffer from limitations in precision and stability. Although deep learning has improved detection accuracy, it increases the demand on computational resources. To address these issues, this paper proposes an optimized lightweight deep learning model called S-GhostNet, which aims to enhance the efficiency and accuracy of magnetic tile surface defect detection while reducing computational complexity. S-GhostNet employs advanced optimization techniques, such as generating Ghost features using different dilation rates in convolutional layers to capture multi-scale defect information, thereby enhancing feature diversity. Channel shuffling and depthwise separable convolutions promote feature fusion and reduce redundant computations. Additionally, the integration of Feature Pyramid Networks (FPN) improves the detection of defects of various sizes. Experimental results show that S-GhostNet achieves an accuracy of 95.46\% in magnetic tile surface defect detection, achieving a 14.36\% improvement over the original GhostNet, while reducing computational cost (FLOPs) by approximately 29.76\%, from 4.20 G to 2.95 G. This demonstrates that S-GhostNet not only enhances detection accuracy but also significantly reduces the required computational resources, highlighting its advantages in this field.