Infrared and Laser Engineering, Volume. 54, Issue 8, 20250209(2025)
Dynamic feature aggregation and multi-level collaboration for UAV infrared target instance segmentation
Figures & Tables(14)
Fig. 1. Dynamic feature aggregation and multilevel collaborative segmentation model
Table 1. Hyperparametric configuration
View tableView in ArticleTable 1. Hyperparametric configuration
Parameter Value Epoch 300 Learning rate 0.01 Batch size 16 Weight decay 0.0005
Table 2. Ablation experiment
View tableView in ArticleTable 2. Ablation experiment
Models SDAConv FARU MSFE GFLOPs Size/MB mAP50 Inference/ms YOLOv8n 12.0 6.8 67.6% 5.5 Ours1 √ 12.7 7.2 71.8% 10.6 Ours2 √ √ 13.0 7.5 74.3% 14.0 DFANet √ √ √ 13.1 7.7 78.4% 10.8
Table 3. Comparative experiments of different backbone networks
View tableView in ArticleTable 3. Comparative experiments of different backbone networks
Models GFLOPs Size/MB mAP50 Inference/ms ShuffleNetv2 14.9 13.2 68.7% 10.9 MobileNetv3 10.9 9.9 67.5% 10.6 GhostNet 6.6 9.6 68.9% 12.1 EfficientNet 7.5 7.1 69.1% 11.8 Ours 12.7 7.2 71.8% 10.6
Table 4. Comparison experiments of different dynamic convolutional networks
View tableView in ArticleTable 4. Comparison experiments of different dynamic convolutional networks
Models GFLOPs Size/MB mAP50 Inference/ms SCConv 12.1 6.8 64.2% 8.4 DySnakeConv 12.1 7.3 66.1% 8.3 ODConv 12.7 6.8 68.8% 8.4 DGConv 12.9 6.8 63.5% 9.5 SADConv 12.7 7.2 71.8% 10.6
Table 5. Comparison experiments of different feature pyramids
View tableView in ArticleTable 5. Comparison experiments of different feature pyramids
Models GFLOPs Size/MB mAP50 Inference/ms SPPF 12.0 6.8 67.6% 5.5 SimSPPF 12.1 6.7 66.5% 3.2 SAPP 11.8 7.8 69.3% 7.1 PPM 13.4 10.8 68.1% 10.7 DAPPM 13.5 11.2 68.9% 12.3 MSFE 12.7 7.2 72.6% 8.1
Table 6. Comparative experiments of different networks
View tableView in ArticleTable 6. Comparative experiments of different networks
Models GFLOPs Size/ MB mAP50 mAP50-95 Inference/ ms YOLACT 13.9 7.4 62.3% 43.4% 11.3 YOLACT++ 13.1 7.7 62.8% 45.1% 11.7 YOLOv5n-seg 11.8 8.3 65.3% 44.2% 2.6 YOLOv7-Tiny 14.9 12.5 65.8% 46.1% 9.8 YOLOv8n-seg 12.0 6.8 67.6% 45.5% 5.5 YOLOv8s-seg 32.4 22.7 77.1% 52.5% 11.6 YOLOv11n-seg 10.2 6.0 72.3% 48.0% 3.7 YOLOv11s-seg 35.3 20.5 76.4% 53.3% 10.9 YOLOv12n-seg 10.2 5.8 71.2% 45.9% 4.8 YOLOv12s-seg 35.2 19.3 76.3% 52.8% 8.5 DFANet 13.1 7.7 78.4% 51.1% 10.8
Table 7.
K -value cross-validation experimentView tableView in ArticleTable 7.
K -value cross-validation experimentK=5 mAP50 mAP95 A 77.5% 50.7% B 78.9% 52.0% C 79.1% 51.8% D 78.1% 50.7% E 79.6% 51.1% Average precision 78.64% 51.26% Standard deviation 0.78% 0.56%
Tools
Get Citation
Copy Citation Text
Zifen HE, Qigang WANG, Yinhui ZHANG, Ying HUANG, Wei PENG, Guangchen CHEN. Dynamic feature aggregation and multi-level collaboration for UAV infrared target instance segmentation[J]. Infrared and Laser Engineering, 2025, 54(8): 20250209
Paper Information
Category: Optical imaging, display and information processing
Received: Mar. 4, 2025
Accepted: --
Published Online: Aug. 29, 2025
The Author Email: Yinhui ZHANG (zhangyinhui@kust.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20