Journal of Infrared and Millimeter Waves, Volume. 43, Issue 6, 859(2024)
Progressive spatio-temporal feature fusion network for infrared small-dim target detection
Figures & Tables(14)
Fig. 1. Progressive spatio-temporal feature fusion network structure：（a）overall architecture of PSTFNet；（b）progressive temporal accumalation module；（c）multi-scale spatial feature fusion module
Fig. 2. Progressive temporal accumulation module：（a）architecture of the P2DConv module；（b）architecture of the M3DConv module
Fig. 3. SHU-MIRST dataset simulation flowchart：（a）background shooting；（b）target template production；（c）target 3D modeling；（d）image fusion algorithm for region resampling；（e）target template embedding
Fig. 4. SHU-MIRST dataset statistical information: (a) distribution of target sizes；(b) distribution of mean SCR
Fig. 5. Examples of target motion trajectory in the SHU-MIRST dataset
Fig. 6. ROC curves of PSTFNet under different mSCR: (a) mSCR≤3；(b) mSCR>3；(c) all sequences
Fig. 7. Qualitative comparison results of PSTFNet and 6 benchmark algorithms on the SHU-MIRST Dataset
Fig. 8. Visualization map of PSTFNet and the backbone network ResUNet at different stage of decoder
Table 1. Quantitative comparison of different algorithms on the SHU-MIRST dataset
View tableView in ArticleTable 1. Quantitative comparison of different algorithms on the SHU-MIRST dataset
方法 SHU-MIRST（mSCR≤3） SHU-MIRST（mSCR>3） SHU-MIRST（all） IoU/（%） Pd/（%） Fa（10-6） IoU/（%） Pd/（%） Fa（10-6） IoU/（%） Pd/（%） Fa（10-6） Top-Hat 0.00 0.83 856.81 2.67 11.17 185.81 0.93 4.45 621.96 IPI 0.19 2.75 80.23 2.72 14.75 57.34 1.08 6.95 72.22 PSTNN 0.00 0.14 122.94 2.41 10.31 129.36 0.84 3.70 125.19 WSLCM 0.45 45.80 4 623.48 5.61 80.22 3 562.33 2.26 57.85 4 252.08 WSNM-STIPI 9.61 53.61 35.95 13.67 66.01 36.35 11.03 57.95 36.09 IMNN-LWEC 0.00 0.00 32.24 0.12 3.96 139.76 0.04 1.38 69.87 ASTTV-NTLA 0.00 0.30 80.29 0.40 5.02 34.67 0.14 1.95 64.34 RDIAN 36.40 52.07 36.46 67.36 84.84 15.40 47.23 63.54 29.09 DNANet 38.74 61.82 39.75 74.19 85.56 10.60 51.14 70.13 29.55 ISNet 36.17 49.01 13.15 65.33 82.46 13.23 46.38 60.72 13.18 UIUNet 43.54 55.93 11.88 74.29 90.61 3.28 54.30 68.07 8.87 SSTNet - 64.09 18.55 - 93.56 8.92 - 74.40 15.17 ResUNet-DTUM 51.78 68.51 13.32 75.53 93.83 6.60 60.09 77.37 10.97 DNANet-DTUM 51.91 69.19 21.63 76.71 93.98 2.67 60.59 77.86 15.00 Ours 57.68 75.80 10.80 76.28 95.08 2.69 64.19 82.55 7.97
Table 2. Quantitative comparison of different algorithms on the IRDST-Real dataset
View tableView in ArticleTable 2. Quantitative comparison of different algorithms on the IRDST-Real dataset
方法 IoU/（%） Pd/（%） Fa（10-6） Top-Hat 5.39 24.66 489.28 IPI 9.38 36.55 37.11 PSTNN 5.79 17.58 57.05 WSLCM 4.92 37.44 1 389.62 WSNM-STIPI 17.79 59.66 38.92 IMNN-LWEC 3.10 7.99 641.05 ASTTV-NTLA 0.27 1.82 395.59 RDIAN 47.69 86.04 3.95 DNANet 50.34 82.57 5.15 ISNet 50.35 82.38 3.86 UIUNet 48.73 81.54 2.70 SSTNet - 85.11 4.83 ResUNet-DTUM 50.31 86.19 2.87 DNANet-DTUM 50.98 87.03 3.62 Ours 53.93 91.25 2.26
Table 3. Results of the ablation experiment for the PSTFNet component modules
View tableView in ArticleTable 3. Results of the ablation experiment for the PSTFNet component modules
方法 IoU/（%） Pd/（%） Fa（10-6） Backbone 37.17 50.41 24.89 Backbone + PTAM 58.10 76.23 12.13 Backbone + MSFM 40.58 56.78 11.08 PSTFNet 64.19 82.55 7.97
Table 4. Results of the PTAM layer ablation experiment
View tableView in ArticleTable 4. Results of the PTAM layer ablation experiment
方法 IoU/（%） Pd/（%） Fa（10-6） PSTFNet w/o PTAM 40.58 56.78 11.08 PSTFNet w/o PTAM L123 43.26 58.95 13.14 PSTFNet w/o PTAM L12 49.04 64.02 9.95 PSTFNet w/o PTAM L1 55.33 72.60 10.29 PSTFNet 64.19 82.55 7.97
Table 5. Results of the PTAM composition ablation experiment
View tableView in ArticleTable 5. Results of the PTAM composition ablation experiment
方法 IoU/（%） Pd/（%） Fa（10-6） PSTFNet w/o PTAM 40.58 56.78 11.08 PSTFNet w/o M3DConv 58.74 74.99 7.17 PSTFNet w/o P2Dconv 53.83 66.51 13.23 PSTFNet 64.19 82.55 7.97
Table 6. Results of the MSFM composition ablation experiment
View tableView in ArticleTable 6. Results of the MSFM composition ablation experiment
方法 IoU/（%） Pd/（%） Fa（10-6） PSTFNet w/o MSFM 58.10 76.23 12.13 PSTFNet w/o MC 60.52 79.16 9.25 PSTFNet w/o SA 62.47 81.38 17.18 PSTFNet 64.19 82.55 7.97
Tools
Get Citation
Copy Citation Text
Dan ZENG, Jian-Ming WEI, Jun-Jie ZHANG, Liang CHANG, Wei HUANG. Progressive spatio-temporal feature fusion network for infrared small-dim target detection[J]. Journal of Infrared and Millimeter Waves, 2024, 43(6): 859
Paper Information
Category: Interdisciplinary Research on Infrared Science
Received: Mar. 24, 2024
Accepted: --
Published Online: Dec. 13, 2024
The Author Email: HUANG Wei (lyxhw@shu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20