Balancing complexity and accuracy for defect detection on filters with an improved RT-DETR.

Filters are critical components in automotive engine systems, responsible for maintaining stable operation by removing impurities from liquids and gases. Their performance is highly sensitive to surface defects, rendering high-precision automated inspection essential. However, existing defect detection algorithms often struggle to balance between detection accuracy and the computational efficiency required for industrial deployment. To address this trade-off, this study introduces an improved detection method based on the Real-Time DEtection TRansformer(RT-DETR) framework. First, a large-kernel attention mechanism is integrated into the backbone to enhance multi-scale feature extraction and fusion, while reducing architectural redundancy. Second, the RepC3 structure within the cross-scale fusion module is replaced with a module based on the generalized-efficient layer aggregation network that uses a more efficient layer aggregation strategy to improve feature localization. Finally, the Adown downsampling module is introduced, employing a multi-path design that reduces parameter count while preserving critical feature details during scale reduction. Experimental results on our industrial filter surface defect dataset show that the enhanced RT-DETR model achieves a mean average precision of 97.6%, a 7.3 percentage point increase over the baseline. Furthermore, the model reduces parameter count by 6.9% and computational load by 13.1%, demonstrating its improved efficiency. Generalization experiments on the public NEU-DET dataset and GC10-DET dataset further confirm the model's robustness and effectiveness, demonstrating its suitability for industrial applications requiring both high accuracy and lightweight deployment.