Real‐Time Gear Surface Defect Detection via Multi‐Order Feature Aggregation and Spatial–Frequency Dual‐Domain Enhancement

Le résumé fourni par la source

Gears present unique inspection challenges due to involute tooth geometries, diverse industrial interferences such as lubricant residues and metallic reflections, and defect scales spanning from micrometer-level cracks to millimeter-level fractures. To address these challenges, this article presents MPC-DETR with improved matching (DEIM), a real-time gear defect detection model that achieves an optimal balance between accuracy and efficiency through multi-order feature aggregation and dual-domain optimization. The architecture employs a multi-order gated aggregation backbone to capture minute defects within complex gear textures while maintaining computational efficiency. Subsequently, a depthwise frequency-spatial convolution module decouples defect signals from background noise via spatial-frequency collaborative processing, capitalizing on the distinct spectral characteristics inherent to industrial interference. Building upon these refined features, an efficient multi-scale fusion module leverages shared large-kernel convolutions with dynamic kernel generation, enabling precise localization across heterogeneous defect scales. Extensive evaluations on a self-constructed GEER-DET dataset demonstrate that MPC-DEIM attains 95.2% mean average precision (mAP)@0.5, surpassing the DEIM baseline by 5.1% while reducing computational costs by 56.6%, and comprehensively outperforming state-of-the-art detection transformer (DETR)-series and you-only-look-once (YOLO)-series algorithms. Validation on the public GSD dataset achieves 98.8% mAP@0.5 with a 5.5 percentage point improvement. Additional cross-dataset evaluations on NEU-DET, PCB-DET, and MS COCO further confirm broad cross-industry applicability.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.