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

Area of Science:

• * Materials Science
• * Mechanical Engineering
• * Computer Vision

Background:

• * Automotive filters require high-precision inspection for surface defects to ensure stable engine operation.
• * Existing automated defect detection algorithms face challenges in balancing accuracy and computational efficiency for industrial use.
• * Surface defects significantly impact filter performance and longevity.

Purpose of the Study:

• * To develop an improved defect detection method for automotive filters.
• * To address the trade-off between detection accuracy and computational efficiency in industrial inspection.
• * To enhance the Real-Time DEtection TRansformer (RT-DETR) framework for filter surface defect analysis.

Main Methods:

• * Integration of a large-kernel attention mechanism into the RT-DETR backbone for enhanced multi-scale feature extraction.
• * Replacement of the RepC3 structure with a generalized-efficient layer aggregation network module for improved feature localization.
• * Introduction of an Adown downsampling module with a multi-path design to preserve feature details during scale reduction.

Main Results:

• * The enhanced RT-DETR model achieved a mean average precision of 97.6% on an industrial filter surface defect dataset, a 7.3% increase over the baseline.
• * Parameter count was reduced by 6.9%, and computational load decreased by 13.1%, indicating improved efficiency.
• * Generalization experiments on NEU-DET and GC10-DET datasets confirmed the model's robustness and effectiveness.

Conclusions:

• * The proposed enhanced RT-DETR model offers a superior solution for automated filter surface defect detection.
• * The method successfully balances high accuracy with lightweight deployment requirements for industrial settings.
• * This approach is suitable for real-world industrial applications demanding efficient and precise defect identification.