MC-DETR: a lightweight RT-DETR method for industrial defect detection

Abstract

Industrial defect detection requires accurate localization of small and irregular defects under complex backgrounds while maintaining low computational cost for deployment-oriented applications. To address the redundancy and limited small-defect representation of RT-DETR in industrial scenarios, this paper proposes MC-DETR, a lightweight detector built on RT-DETR. The method introduces an MSInit-cross stage partial network backbone for multi-scale lightweight feature extraction, a curvature-guided token attention module for defect-relevant token interaction, and an NWDIoU-based regression strategy for more stable localization. Experiments on a self-built industrial defect dataset and the public NEU-DET dataset show that MC-DETR reduces parameters by 34.5% and GFLOPs by 21.1% compared with RT-DETR-R18, while achieving mAP0.5 of 0.803 and mAP0.5:0.95 of 0.458 on the self-built dataset. The results indicate that MC-DETR provides a favorable accuracy-complexity trade-off for industrial defect inspection.