Chinese Optics, Volume. 17, Issue 3, 538(2024)
Indistinguishable points attention-aware network for infrared small object detection
Fig. 1. Indistinguishable points attention-aware network architecture
Fig. 2. (a) Centre point offset to boundary; (b) point-based region proposal module
Fig. 3. ROC curves for methods on (a) the NUDT-SIRST dataset and (b) the IRDST dataset
Fig. 4. Different scene detection results of different methods, with the same target zoomed in the same color box
Fig. 6. Visualization of regional proposals. (a) Original maps; (b) groundtruths; (c) centroid heatmap and regional proposal boundaries
Fig. 7. Visualization of indistinguishable points distribution. (a) UAV target; (b) point target; (c) aircraft target
Table 1. Hyperparameter settings of traditional algorithms
View tableView in ArticleTable 1. Hyperparameter settings of traditional algorithms
传统算法 超参数设置 Top-hat Nhood=ones(5) LEF h=0.2，α=0.5， P=9 AADCDD 内窗口尺寸={3, 5, 7, 9}，外窗口尺寸=19 TLLCM 窗口尺寸={3, 5, 7, 9}，k=9
Table 2. Comparison of quantitative results of different methods on NUDT-SIRST and IRDST datasets
View tableView in ArticleTable 2. Comparison of quantitative results of different methods on NUDT-SIRST and IRDST datasets
检测算法 NUDT-SIRST IRDST mAP F值（Pre，Rec） mAP F值（Pre，Rec） Top-hat 1.5 0.3599（0.2850，0.4884） 0.7 0.0088（0.0045，0.4107） LEF 6.4 0.1151（0.0748，0.2498） 2.5 0.1219（0.0686，0.5470） AADCDD 1.6 0.1490（0.3838，0.0924） 1.4 0.0705（0.0521，0.1090） TLLCM 16.5 0.0724（0.0479，0.1476） 6.1 0.1881（0.1254，0.3759） ALCNet 69.3 0.7595（0.7035，0.8251） 46.5 0.5929（0.5461，0.6486） DNANet 86.9 0.8645（ 0.9070 ，0.8259）62.1 0.6697（ 0.712 4 ，0.6319）RDIAN 82.4 0.890 0 （0.899 0 ，0.881 1 ）60.0 0.7102 （0.7092，0.7113 ）本文方法 87.4 0.8935 （0.8923，0.8948 ）63.4 0.7056（ 0.7183 ，0.6935）
Table 3. Average inference times of a single image for deep learning methods
(s) View tableView in ArticleTable 3. Average inference times of a single image for deep learning methods
(s) 检测算法 NUDT-SIRST IRDST ALCNet 0.104 0.166 DNANet 0.089 0.259 RDIAN 0.065 0.114 本文算法 0.099 0.121
Table 4. Comparison of different region proposal modules
View tableView in ArticleTable 4. Comparison of different region proposal modules
建议数量 基于点的区域建议 RPN mAP F值 mAP F值 1000 87.9 0.8927 86.2 0.8425 256 87.5 0.8962 85.8 0.8412 128 87.4 0.8935 85.2 0.8406 64 86.0 0.8901 84.5 0.8397
Table 5. Detection results of different point selection strategies
View tableView in ArticleTable 5. Detection results of different point selection strategies
选点策略 mAP 均匀选点 86.7 k=1,γ=0.00 86.9 k=3,γ=0.75 87.4 k=10,γ=1.00 85.8
Table 6. Fusion results of different features at indistinguishable points
View tableView in ArticleTable 6. Fusion results of different features at indistinguishable points
细粒度特征 粗糙掩码 位置嵌入 mAP √ 85.5 √ √ 85.8 √ √ √ 87.4
Table 7. Results of different refinement strategies
View tableView in ArticleTable 7. Results of different refinement strategies
细化方案 mAP CNN（16×16） 85.5 MLP（16×16） 86.2 细化掩码边界模块（S=3） 87.4 细化掩码边界模块（S=6） 87.6
Get Citation
Copy Citation Text
Bo-xiao WANG, Yan-song SONG, Xiao-na DONG. Indistinguishable points attention-aware network for infrared small object detection[J]. Chinese Optics, 2024, 17(3): 538
Category: Original Article
Received: Oct. 11, 2023
Accepted: Dec. 5, 2023
Published Online: Jul. 31, 2024
The Author Email:
