Authors
Chenyang Li, Jie Cao, Qun Hao, Chenghao Song, Haifeng Yao, Zhipeng Wei
Published in
Sensors (Basel, Switzerland). Volume 26. Issue 15. Aug 01, 2026. Epub Aug 01, 2026.
Abstract
Infrared multi-class small-target detection is challenging because targets occupy few pixels, exhibit weak texture, and are easily confused with background clutter. We present RADFRNet, a YOLOv11-n-based detector designed to address two forms of information degradation: detail loss in the backbone and semantic-spatial mismatch during cross-level feature fusion. First, the previously proposed DEConv operator is embedded into selected C3K2 stages to form C3DEConv; the contribution lies in its C3K2-compatible, detector-oriented integration rather than in a new differential-convolution formulation. Second, an Adaptive Feature Recalibration (AFRE) block constructed from three Recalibration Attention Units performs bidirectional interaction between shallow spatial details and deep semantic features. We also construct four-class bounding-box annotations for BIT-SIRST. RADFRNet achieves [email protected] scores of 93.2% and 65.3% on BIT-SIRST and FLIR-ADAS-v2, improving YOLOv11-n by 4.4 and 8.5 percentage points, respectively. Under the same original 640×640 GPU inference setup, the per-image latency increases from 3.3 to 6.7 ms on BIT-SIRST and from 3.0 to 7.8 ms on FLIR-ADAS-v2, corresponding to nominal throughputs of approximately 149 and 128 FPS for RADFRNet. The reported model-complexity values are 8.2 M and 8.6 M, respectively. These results show that RADFRNet retains high-rate GPU inference, although the accuracy gains are obtained at a clear computational cost; the model is therefore positioned as an accuracy-oriented detector rather than a latency-neutral replacement for YOLOv11-n.
PMID:
42590627
Bibliographic data and abstract were imported from PubMed on 13 Aug 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 5
- Comments 0