Improved UAV Target Detection Model for RT-DETR
In light of the shortcomings pertaining to UAV small target detection, the detection of complex scenes, and the detection of multi-scale targets, a time-frequency dual-domain feature extraction algorithm, TF-DETR, has been proposed. This algorithm has been optimized for RT-DETR.Firstly, a time-frequ...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11018325/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In light of the shortcomings pertaining to UAV small target detection, the detection of complex scenes, and the detection of multi-scale targets, a time-frequency dual-domain feature extraction algorithm, TF-DETR, has been proposed. This algorithm has been optimized for RT-DETR.Firstly, a time-frequency domain feature extraction module, TF-CSPNet, has been introduced into the backbone network. This module facilitates the efficient extraction and fusion of multi-source features. Secondly, the Extreme Perceptive Linear Attention (EPLA) mechanism is designed and introduced to improve the AIFI module, which enhances the model’s attention to the key information by considering the positive and negative polarity interactions between the query and the key. Furthermore, the Focaler-MPDIoU loss function has been developed to address the challenge of suboptimal localization accuracy for hard-to-detect targets and diminutive targets. On the VisDrone2019 dataset, the mAP0.5 of the enhanced model demonstrates a 3.5% improvement, accompanied by a 6.1% and 2.9% reduction in parameters and computations, respectively. The efficacy of these enhancements is substantiated by the model’s superior performance in comparison to other target detection models at equivalent levels. |
|---|---|
| ISSN: | 2169-3536 |