Lightweight RT-DETR with small-object feature enhancement for traffic sign detection
Traffic sign detection is safety-critical for vision-based autonomous driving, yet difficult for two coupled reasons: signs span very few pixels at the distance a vehicle must first react, and detectors accurate enough to resolve them consume much of the compute and memory that an on-vehicle system shares among several networks. The real-time detection transformer removes the nonmaximum-suppression bottleneck but remains parameter-heavy and weak on small, distant signs. We address weight and small-object accuracy jointly through three components. RepGhost–CSP-ELAN replaces the backbone, suppressing feature redundancy with cheap linear operations, whereas re-parameterisable convolutions restore gradient flow at no deployment cost. A small-object feature-enhancement pyramid injects a high-resolution P2 branch through a space-to-depth convolution that preserves fine detail, fused by a CSP-partitioned OmniKernel block at bounded cost. Focaler-GIoU-NWD, a regression loss coupling Focaler-GIoU with a normalized Wasserstein distance, stays stable on boxes only a few pixels across. On TT100K, the model raises [email protected] from 82.9% to 89.5% while reducing parameters by 27.1% and raising throughput by 8.5%, outperforming all evaluated YOLO detectors. The gain concentrates where it matters: small-object APS rises from 47.2% to 60.3%, roughly twice the overall gain. A lighter variant reaches 87.3% [email protected] with 10.5 M parameters, a second accuracy–efficiency operating point.
Authors
- Bohang Chao (ORCID: https://orcid.org/0009-0000-1682-4790)
- You Zhou (ORCID: https://orcid.org/0009-0004-4590-2638)
Institutions
- Donghua University (CN)
Publication Details
- Journal
- Journal of Electronic Imaging
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1117/1.jei.35.6.063011
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00