Laser & Optoelectronics Progress, Volume. 62, Issue 14, 1439003(2025)
Infrared-Visible Image Fusion Network Based on Dual-Branch Feature Decomposition
Figures & Tables(15)
Fig. 1. Network framework of DBDFuse. (a) Flowchart in training phase; (b) fusion layer in testing phase
Fig. 4. Visualization of SOTA methods for infrared and visible light image fusion
Table 1. Quantitative results of IVIF task on the LLVIP dataset
View tableTable 1. Quantitative results of IVIF task on the LLVIP dataset
Method EN SF SD MI SCD VIF DIDFuse 7.05 19.2 48.57 1.82 1.43 0.68 DenseFuse 7.00 16.10 50.21 1.87 1.32 0.64 CDDFuse 7.34 20.54 52.40 2.16 1.68 0.78 FusionGan 6.45 12.04 39.4 1.71 1.54 0.51 ReCoNet 7.35 14.63 44.89 1.48 1.23 0.62 TarDAL 6.51 10.65 53.12 1.56 1.46 0.63 U2Fusion 6.78 19.70 48.02 1.89 1.51 0.59 DBDFuse 7.42 23.36 59.18 2.56 1.52 0.80
Table 2. Quantitative results of IVIF task on the TNO dataset
View tableTable 2. Quantitative results of IVIF task on the TNO dataset
Method EN SF SD MI SCD VIF DIDFuse 7.03 13.12 42.66 1.71 1.73 0.63 DenseFuse 6.98 8.79 39.89 1.78 1.64 0.66 CDDFuse 7.12 13.17 46.53 2.09 1.76 0.71 FusionGan 6.67 8.77 30.03 1.63 1.68 0.34 ReCoNet 7.10 8.75 44.87 1.79 1.70 0.58 TarDAL 6.87 8.68 46.63 1.86 1.56 0.53 U2Fusion 6.86 11.58 35.57 1.47 1.71 0.59 DBDFuse 7.15 13.23 48.10 1.96 1.79 0.72
Table 3. Quantitative results of IVIF task on the MSRS dataset
View tableTable 3. Quantitative results of IVIF task on the MSRS dataset
Method EN SF SD MI SCD VIF DIDFuse 5.64 10.16 32.49 1.61 1.21 0.41 DenseFuse 6.25 8.60 38.25 2.15 1.36 0.78 CDDFuse 6.71 11.54 43.36 3.46 1.62 1.07 FusionGan 5.89 8.70 34.38 2.61 1.08 0.24 ReCoNet 6.72 11.02 43.44 2.16 1.45 0.74 TarDAL 5.65 8.25 26.30 1.49 0.72 0.45 U2Fusion 6.24 9.08 25.51 1.40 1.32 0.55 DBDFuse 6.89 11.98 43.64 2.98 1.64 1.12
Table 4. Ablation experimental results on the LLVIP dataset
View tableTable 4. Ablation experimental results on the LLVIP dataset
Method EN SF SD SCD VIF Ⅰ：+Outlook 7.06 22.62 58.64 1.48 0.70 Ⅱ： +Stoken 6.94 21.43 57.43 1.50 0.76 Ⅲ： +CNN 6.53 20.57 50.67 1.34 0.56 Ⅳ： +Outlook+CNN 6.82 20.86 52.35 1.38 0.61 Ⅴ： +Stoken+CNN 6.21 20.96 54.5 1.40 0.71 DBDFuse 7.42 23.36 59.18 1.52 0.80
Table 5. Quantitative comparison of the proposed method with different loss functions
View tableTable 5. Quantitative comparison of the proposed method with different loss functions
Configuration EN SF SD SCD VIF L1 6.09 20.22 50.82 1.35 0.66 L2 5.94 18.48 52.54 1.15 0.57 DBDFuse 7.42 23.36 59.18 1.52 0.80
Table 6. Overall fusion time of 50 pairs of images
View tableTable 6. Overall fusion time of 50 pairs of images
Method Image pair fusion time /s TNO （25） LLVIP （25） MSRS （25） DenseFuse 70.68 76.20 72.41 DIDFuse 39.49 45.25 40.57 CDDFuse 32.58 42.64 36.52 FusionGan 150.45 175.74 169.49 ReCoNet 70.59 74.23 70.89 TarDAL 50.7 59.66 52.12 U2Fusion 65.68 94.70 69.02 DBDFuse 33.31 42.36 33.18
Table 7. Target detection evaluation value on the M3FD dataset
View tableTable 7. Target detection evaluation value on the M3FD dataset
Method Car Lamp People Truck Bus Motorcycle mAP@0.5 VI 90.46 87.52 70.95 70.78 78.23 70.34 78.04 IR 88.65 71.68 80.76 66.2 78.62 64.23 75.02 DenseFuse 91.87 84.12 79.86 69.84 81.65 67.98 79.22 DIDFuse 92.75 87.84 80.13 71.54 82.98 69.24 80.74 CDDFuse 92.58 82.78 81.77 71.56 82.62 71.75 80.51 FusionGan 92.74 84.65 79.77 69.44 79.55 68.77 79.15 ReCoNet 92.44 87.62 80.46 71.65 79.84 67.45 79.91 TarDAL 94.75 87.11 81.12 68.45 81.33 69.45 80.36 U2Fusion 91.45 88.12 79.24 69.54 78.92 69.98 79.54 DBDFuse 92.65 86.74 81.75 71.58 82.64 70.56 80.98
Table 8. Semantic segmentation evaluation values on the LLVIP dataset.
View tableTable 8. Semantic segmentation evaluation values on the LLVIP dataset.
Method Car Guardrail Pedestrian Obstacle Color cone Bike Car stop Background mIOU VI 75.4 46.1 45.8 50.1 43.2 59.2 51.8 90.2 55.4 IR 67.4 42.9 46.2 48.2 40.2 51.8 35.2 88.3 50.6 DenseFuse 79.4 61.7 58.6 56.4 42.6 61.5 49.8 97.1 60.4 DIDFuse 78.6 51.2 63.0 59.8 44.6 61.2 52.4 97.5 60.9 CDDFuse 82.4 58.7 62.7 51.6 44.1 63.7 53.8 97.6 61.2 FusionGan 78.9 48.6 60.1 49.6 47.8 60.9 48.4 96.8 58.6 ReCoNet 80.2 60.1 55.7 53.7 42.8 62.7 49.5 97.5 59.5 TarDAL 79.6 59.4 61.2 54.8 46.3 64.4 50.3 97.1 61.5 U2Fusion 79.5 57.6 59.7 55.7 47.2 63.9 51.7 97.4 61.6 DBDFuse 83.1 62.5 64.0 57.1 49.6 65.4 53.1 97.7 63.9
Tools
Get Citation
Copy Citation Text
Xundong Gao, Hui Chen, Yaning Yao, Chengcheng Zhang. Infrared-Visible Image Fusion Network Based on Dual-Branch Feature Decomposition[J]. Laser & Optoelectronics Progress, 2025, 62(14): 1439003
Paper Information
Category: AI for Optics
Received: Dec. 23, 2024
Accepted: Mar. 2, 2025
Published Online: Jul. 16, 2025
The Author Email: Xundong Gao (xundonggao@guet.edu.cn), Hui Chen (Chenhui02@guet.edu.cn)
CSTR:32186.14.LOP242481
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20