Infrared and Laser Engineering, Volume. 54, Issue 7, 20240598(2025)
Uav infrared object detection algorithm based on multi-scale fusion and channel compression
Figures & Tables(18)
Fig. 5. GSConv module including splicing structure and channel scrubbing structure
Fig. 9. (a) YOLOv8n baseline method confusion matrix; (b) Improved method confusion matrix
Table 1. Exprimental parameter configuration
View tableView in ArticleTable 1. Exprimental parameter configuration
Parameter Value Optimizer SGD Epochs/epochs 300 Lr0 1e-2 Images size/pixel 640×640 Momentum 0.937 Patience/epochs 10
Table 2. Ablation experiment results on the HIT-UAV validation set
View tableView in ArticleTable 2. Ablation experiment results on the HIT-UAV validation set
Methods mAP@0.5 mAP@0.5:0.95 Params/M FLOPs/G Module size/M FPS/fps HGNetv2 PSA Slim-neck Inner-EIou × × × × 89.0% 59.5% 3.1 8.2 6.3 131.7 √ × × × 87.9% 59.0% 2.3 6.6 4.6 143.9 √ √ × × 90.2% 60.4% 2.3 6.7 5.0 138.6 √ √ √ × 90.7% 60.6% 2.0 5.9 4.7 152.9 √ √ √ √ 90.5% 61.4% 2.0 5.9 4.7 152.9
Table 3. Comparison of lightweight backbone networks
View tableView in ArticleTable 3. Comparison of lightweight backbone networks
Backbone mAP@0.5 Params/M FLOPs/G Module size/M CSPDarkNet 89.0% 3.1 8.2 6.3 MobileNetV4 85.6% 4.3 8.1 8.5 Ghostnetv2 89.9% 6.3 8.7 11.2 ShuffleNetV2 88.7% 2.7 7.5 5.9 HGNetV2 87.9% 2.3 6.6 4.6 PHGNet 90.2% 2.3 6.7 5.0
Table 4. Comparison of loss functions
View tableView in ArticleTable 4. Comparison of loss functions
IoU mAP@0.5 mAP@0.5:0.95 CIoU(baseline) 90.7% 60.6% EIoU 88.5% 60.3% FocusIoU 89.4% 59.9% Inner-CIoU 89.7% 61.0% Inner-EIoU 90.5% 61.4%
Table 5. Comparison of experimental results with eifferent model algorithms
View tableView in ArticleTable 5. Comparison of experimental results with eifferent model algorithms
Models mAP@0.5 mAP@0.5:0.95 Params/M FLOPs/G Module size/M FPS/fps Faster-RCNN 81.1% 53.1% 137.0 310.5 109.2 69.0 YOLOv4-tiny 78.2% 51.9% 6.1 6.9 5.9 136.3 YOLOv5s 90.8% 60.4% 7.2 16.3 14.4 146.2 YOLOv6n 87.0% 58.3% 4.2 11.9 8.7 149.2 YOLOv7-tiny 89.6% 59.3% 6.0 13.2 11.7 128.9 YOLOv10n 87.0% 58.0% 2.7 6.5 5.8 140.6 YOLOv8n 89.0% 59.5% 3.1 8.2 6.3 131.7 PSI-YOLO 90.5% 61.4% 2.0 5.9 4.7 152.9
Tools
Get Citation
Copy Citation Text
Kaijun WU, Zhibo WAN, Juanjuan DU, Lidong ZHANG, Yuelian WU, Fengqi ZHANG. Uav infrared object detection algorithm based on multi-scale fusion and channel compression[J]. Infrared and Laser Engineering, 2025, 54(7): 20240598
Paper Information
Category: Optical imaging, display and information processing
Received: Dec. 23, 2024
Accepted: --
Published Online: Aug. 29, 2025
The Author Email: Zhibo WAN (12231965@stu.lzjtu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20