Steel surface defect detection is a crucial step in improving product quality, ensuring safety, and enhancing production efficiency. However, existing object detection models still face challenges such as high miss rates and computational complexity when handling diverse defect types and complex background noise. To address these issues, this study proposes a steel surface defect detection model based on the real-time detection transformer, aimed at improving detection accuracy in complex industrial environments. To tackle the challenge of significant shape variations among defects, a dynamic scale aggregation module is designed to fuse critical information from multiple pathways, enabling efficient multi-scale feature extraction. Additionally, a frequency-enhanced dynamic attention mechanism is proposed, utilising the fast Fourier transform for global context modelling in the frequency domain, enhancing the joint perception of low-frequency structural and high-frequency detailed features. Finally, a wavelet upsampling operator is developed to decompose input features into high- and low-frequency components, enabling targeted refinement and adaptive fusion. Experimental results show that the improved model achieves AP of 46.6% and 36.0% on the NEU-DET and GC10-DET datasets, respectively, surpassing the baseline by 2.7% and 1.9%, while reducing FLOPs by 10.0 G. This contributes to more reliable surface quality control in steel processing workflows.
Gan et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: