DRP-RTDETR-based surface defect detection for steel wire ropes

To address the challenges of low detection accuracy rates due to complex textures, tiny defect scales, and industrial scene interference in steel wire rope surface defect detection, this paper proposes a novel algorithm based on DRP-RTDETR. In order to enhance the model's ability to identify complex textures and minute defects, a high-and-low frequency feature fusion module (PSFM) is integrated into the neck network, leveraging cross-attention mechanisms to consolidate multilevel feature information. Simultaneously, the DRBC3 module supersedes the RepC3 module within the hybrid encoder. By combining large kernel and parallel small kernel dilated convolutions, it effectively fuses multi-scale information without incurring additional inference costs. Experimental results demonstrate that, compared to RT-DETR, the DRPRTDETR model achieves a 4.1% improvement in accuracy and a 2.5% enhancement in average precision, reaching 91.2% and 85.5%, respectively. Additionally, Params and FLOPs are reduced by 8.54% and 4.91%. However, despite these improvements, the algorithm's generalization ability requires further refinement. Future research will concentrate on expanding the dataset, diversifying defect categories, and investigating deployment strategies on edge computing devices. The ultimate goal is to construct a more efficient and reliable steel wire rope defect detection system.