Research Article
DPS-DETR: Lightweight Transformer-Based Object Detection for Industrial Safety Monitoring
@INPROCEEDINGS{10.1007/978-3-032-33662-0_38, author={YunTao Zhao and FengYi Wu}, title={DPS-DETR: Lightweight Transformer-Based Object Detection for Industrial Safety Monitoring}, proceedings={Communications and Networking. 20th EAI International Conference, ChinaCom 2025, Xi’an, China, November 8--9, 2025, Proceedings}, proceedings_a={CHINACOM}, year={2026}, month={10}, keywords={Power plant safety monitoring； DPS-DETR； Deformable Attention(DAttention)；Scale-aware Dynamic Loss}, doi={10.1007/978-3-032-33662-0_38} }- YunTao Zhao
FengYi Wu
Year: 2026
DPS-DETR: Lightweight Transformer-Based Object Detection for Industrial Safety Monitoring
CHINACOM
Springer
DOI: 10.1007/978-3-032-33662-0_38
Abstract
To address the challenges of low small-target detection accuracy, limited robustness, and inadequate real-time performance in complex industrial environments, this study proposes DPS-DETR, an enhanced RT-DETR framework. The model integrates three key improvements: Deformable Attention to strengthen feature representation of irregularly distributed small and distant targets, Partial Convolution (PConv) for enhanced spatial detail retention during downsampling, and a Scale-aware Dynamic Loss (SD Loss) embedded in a PSConv module to adaptively reweight losses in favor of small-target localization. Experimental evaluations demonstrate that DPS-DETR achieves 63.04% mAP (+1.63% over RT-DETR) with 54.5 GFLOPs, effectively balancing accuracy and computational efficiency. Moreover, the model exhibits superior robustness and detection granularity under severe occlusion and dense small-target scenarios, underscoring its applicability to intelligent safety supervision in industrial contexts.