FECI-RTDETR a Lightweight Unmanned Aerial Vehicle Infrared Small Target Detector Algorithm Based on RT-DETR
Addressing the challenges of small target detection in aerial infrared images from a drone’s perspective, such as diverse target scales, complex backgrounds, the clustering of small targets, and limited computational resources of the drone platform. This paper proposes a lightweight UAV i...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10836685/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Addressing the challenges of small target detection in aerial infrared images from a drone’s perspective, such as diverse target scales, complex backgrounds, the clustering of small targets, and limited computational resources of the drone platform. This paper proposes a lightweight UAV infrared small target detection algorithm, FECI-RTDETR. Initially, we introduce a lightweight RFConv-Block module that enhances spatial feature extraction capabilities while reducing computational redundancy. Subsequently, we combine the Efficient Additive feature selection mechanism with an intra-scale feature interaction module to form the EA-AIFI module, which strengthens the model’s focus on dense targets and reduces computational burden. Moreover, we introduce the CHS-FPN structure as a cross-scale feature fusion structure, utilizing the coordinate attention mechanism combined with a hierarchical scale-based feature pyramid network. This allows the model to better understand the contextual semantics of targets and improves detection accuracy. Finally, the original GIoU loss is replaced with Inner-GIoU loss, using a scaling factor to control the auxiliary enclosing box, which accelerates convergence speed and enhances detection accuracy for small targets. Experimental results indicate that compared to RT-DETR, the FECI-RTDETR model reduces the number of parameters by 24.56% and floating-point operations by 19.12% on the HIT-UAV dataset. The mAP50 and mAP50:95 metrics improved by 4.2% and 2.9%, respectively, with the mAP50 reaching 84.2%. This algorithmic model achieves a balance between resource reduction and accuracy enhancement while maintaining lightweight characteristics. |
|---|---|
| ISSN: | 2169-3536 |