Chinese Journal of Liquid Crystals and Displays, Volume. 38, Issue 11, 1455(2023)
Infrared dim-small target detection under complex background based on attention mechanism
Fig. 3. Schematic diagram of the attention mechanism.（a）Channel attention module；（b）Spatial attention module；（c）SimAM attention module.
Fig. 4. Structure diagram of network after SimAMC3 module is introduced
Fig. 6. Optimized infrared dim-small target detection network model
Fig. 8. Sample images of typical scenes in NUAA-SIRST.（a）Cloud background；（b）City background；（c）Sea background.
Fig. 9. Sample images of typical scenes in NUDT-SIRST.（a）Cloud background；（b）City background；（c）Sea background；（d）Field background；（e）Highlight background.
Fig. 11. Comparison of test results.（a）IOU inhibition criteria；（b）DIOU inhibition criteria.
Fig. 12. Experimental effect of traditional algorithm.（a）Original image with labelled targets；（b）Top Hat；（c）MLCM；（d）IPI.
Fig. 13. Training model parameter.（a）Mean average precision；（b）Precision；（c）Recall.
Fig. 14. Visual test effect diagram of each network model.（a）Faster R-CNN；（b）SSD；（c）YOLOv3；（d）YOLOv5；（e）Ours；（f）Original image with labelled targets.
Table 1. Main characteristics of the data set selected in this paper
View tableView in ArticleTable 1. Main characteristics of the data set selected in this paper
数据集 数据类型 图像描述 图像数量 NUAA-SIRST 真实红外图像 背景：云层、城市、海面；
目标：亮度微弱，有强光源干扰
427 NUDT-SIRST 合成红外图像 背景：云层、城市、海面、旷野、高亮场景；
目标：亮度微弱，有强光源干扰
1 327
Table 2. Sample classification
View tableView in ArticleTable 2. Sample classification
正类 负类 被检测到 True Positives（TP） False Positives（FP） 未被检测到 False Negatives（FN） True Negatives（TN）
Table 3. Infrared dim-small target data set classification
View tableView in ArticleTable 3. Infrared dim-small target data set classification
Dataset Total Training Val Test NUAA-SIRST 427 257 85 85 NUDT-SIRST 1 327 797 265 265
Table 4. Experimental configuration
View tableView in ArticleTable 4. Experimental configuration
Parameter Configuration OS Ubuntu 20.04 Processor Intel（R）Core（TM）i7-9700K RAM 64.0 GB GPU NVIDIA TITAN RTX GPU Acceleration CUDA 11.2
Table 5. Ablation experiments
View tableView in ArticleTable 5. Ablation experiments
Group SimAMC3 Add Head DIOU_NMS NUAA-SIRST NUDT-SIRST P/% R/% mAP@0.5 P/% R/% mAP@0.5 1 88.7 88.8 91.2 93.7 92.1 94.3 2 √ 91.7 90.8 93.1 95.6 93.9 95.9 3 √ √ 95.8 97.9 97.8 97.2 96.4 98.6 4 √ √ √ 96.9 97.7 98.3 97.9 97.1 99.1
Table 6. Contrast experiment
View tableView in ArticleTable 6. Contrast experiment
Network NUAA-SIRST NUDT-SIRST P/% R/% mAP@0.5/% P/% R/% mAP@0.5/% Faster R-CNN 33.6 46.4 36.3 39.7 49.2 43.6 SSD 62.8 54.3 66.3 70.5 66.1 72.6 YOLOv3 78.9 71.2 82.6 82.9 80.6 85.0 YOLOv5 88.7 88.8 91.2 93.7 92.1 94.3 Ours 96.9 97.7 98.3 97.9 97.1 99.1
Get Citation
Copy Citation Text
Ying LIU, Hai-jiang SUN, Yong-xian ZHAO. Infrared dim-small target detection under complex background based on attention mechanism[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(11): 1455
Category: Research Articles
Received: Jan. 6, 2023
Accepted: --
Published Online: Nov. 29, 2023
The Author Email: Hai-jiang SUN (sunhaijiang@126.com)
