Journal of Infrared and Millimeter Waves, Volume. 41, Issue 6, 1102(2022)
Light-weight infrared small target detection combining cross-scale feature fusion with bottleneck attention module
Figures & Tables(15)
Fig. 2. An illustration of the proposed light-weighted infrared small target detection network
Fig. 5. Samples of eight connected neighborhood clustering module. If the eight neighborhoods of two candidate points have intersection area，they are identified as the same target ID.
Fig. 6. Examples of（a）original images and corresponding qualitative comparison results on（b）Tophat，（c）IPI，（d）RIPT，（e）ACM，（f）DNANet，（g）LIRDNet，（h）ground truth masks.
Fig. 7. Examples of（a）original images and corresponding 3D visualization results on（b）Tophat，（c）IPI，（d）RIPT，（e）ACM，（f）DNANet，（g）LIRDNet，（h）ground truth masks.
Fig. 8. The ROC curve of our proposed LIRDNet under different signal-clutter-ratio（SCR）values（a）SCR<3，（b）3
Fig. 9. Visualization map of our proposed LIRDNet and backbone network ResUnet on different convolutional layers
Table 1. Number of model parameters for LIRDNet and the corresponding size of input and output feature map
View tableView in ArticleTable 1. Number of model parameters for LIRDNet and the corresponding size of input and output feature map
编号 滤波器数量 输入尺寸 输出尺寸 预处理 - 1，256，256 3，256，256 F’de0，0 8 3，256，256 8，256，256 BAM0，0 - 8，256，256 8，256，256 F’de1，0 16 8，128，128 16，128，128 BAM1，0 - 16，128，128 16，128，128 F’de2，0 32 16，64，64 32，64，64 BAM2，0 - 32，64，64 32，64，64 F’de3，0 64 32，32，32 64，32，32 F’de2，1 32 64，64，64 32，64，64 F’de1，1 16 32，128，128 16，128，128 F’de0，1 8 16，256，256 8，256，256 F’final 1 8，256，256 1，256，256
Table 2. Performance of different methods on IoU, Pd, and Fa
View tableView in ArticleTable 2. Performance of different methods on IoU, Pd, and Fa
方法 NUAA-SIRST（条件1） NUAA-SIRST（条件2） IoU/（%） Pd/（%） Fa（10-6） IoU/（%） Pd/（%） Fa（10-6） Top-Hat 17.66 82.56 34.95 7.143 79.84 1012 Max-Median 3.90 52.29 49.32 4.172 69.20 55.33 TLLCM 0.96 77.98 5829 1.029 79.09 5899 IPI 22.77 86.23 10.65 25.67 85.55 11.47 RIPT 11.24 77.98 17.03 11.05 79.08 22.61 MDvsFA-CGAN 63.26 90.75 49.33 60.30 89.35 56.35 ACM 71.78 96.33 3.570 70.33 93.91 3.728 ALCNet 74.39 97.16 22.77 73.33 96.57 30.47 DNANet-Light 74.46 98.19 15.79 74.72 96.95 18.18 LIRDNet-ResNet10 72.52 97.24 22.17 73.47 97.71 26.23 LIRDNet-ResNet18 76.47 98.02 16.85 74.89 97.33 16.09 LIRDNet-ResNet34 77.81 99.21 1.240 75.18 97.33 7.060
Table 3. Performance of different deep learning-based methods on the number of model parameters, FLOPs, IoU, Pd, and Fa
View tableView in ArticleTable 3. Performance of different deep learning-based methods on the number of model parameters, FLOPs, IoU, Pd, and Fa
方法 #Params FLOPs mIoU/Pd/Fa ACM 0.52 M 1.75 G 71.78/96.33/3.570 ALCNet 0.50 M 1.48 G 74.39/97.16/22.77 MDvsFA-cGAN 3.76 M 868.75 G 63.26/90.75/49.33 DNANet-Light 0.48 M 1.88 G 74.46/98.19/15.79 LIRDNet-Res18 0.25 M 1.43 G 76.47/98.02/16.85
Table 4. Inference time and FLOPs performance of different deep learning-based methods on different computational units (Smart Phone-Chip, PC-GPU)
View tableView in ArticleTable 4. Inference time and FLOPs performance of different deep learning-based methods on different computational units (Smart Phone-Chip, PC-GPU)
方法 计算单元 推理时间 /s 浮点运算量 /G DNANet-Light 天玑800U 0.322 1.88 G DNANet-Light 麒麟980 0.211 1.88 G DNANet-Light Nvidia 1070 0.102 1.88 G DNANet-Light Nvidia 3090 0.005 1.88 G LIRDNet 天玑800U 0.198 1.43 G LIRDNet 麒麟980 0.097 1.43 G LIRDNet Nvidia 1070 0.076 1.43 G LIRDNet Nvidia 3090 0.002 1.43 G
Table 5. Ablation study on our proposed CFM module
View tableView in ArticleTable 5. Ablation study on our proposed CFM module
Method #Params FLOPs mIoU/Pd/Fa LIRDNet-Res18
w/o CFM
0.232M 1.184G 73.01/96.58/24.13 LIRDNet-Res18
w/o CFM L1/L2
0.234M 1.204G 73.39/97.16/34.59 LIRDNet-Res18
w/o CFM L1
0.243M 1.362G 74.23/97.16/24.37 LIRDNet-Res18 0.248M 1.435 G 76.47/98.02/16.85
Table 6. Ablation study on our introduced BAM module
View tableView in ArticleTable 6. Ablation study on our introduced BAM module
Method #Params FLOPs mIoU/Pd/Fa LIRDNet-Res18
w/o BAM
0.245M 1.415 G 74.52/96.58/21.29 LIRDNet-Res18
w/o BAM SA
0.247M 1.422 G 75.37/97.16/16.14 LIRDNet-Res18
w/o BAM CA
0.247M 1.434 G 75.43/97.24/25.54 LIRDNet-Res18 0.248M 1.435 G 76.47/98.02/16.85
Tools
Get Citation
Copy Citation Text
Zai-Ping LIN, Bo-Yang LI, Miao LI, Long-Guang WANG, Tian-Hao WU, Yi-Hang LUO, Chao XIAO, Ruo-Jing LI, Wei An. Light-weight infrared small target detection combining cross-scale feature fusion with bottleneck attention module[J]. Journal of Infrared and Millimeter Waves, 2022, 41(6): 1102
Paper Information
Category: Research Articles
Received: Jun. 13, 2022
Accepted: --
Published Online: Feb. 6, 2023
The Author Email: Zai-Ping LIN (linzaiping@sina.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20