Acta Photonica Sinica, Volume. 54, Issue 5, 0510002(2025)
Combining Generative Adversarial Network and Contrastive Learning for Infrared Image Generation
Figures & Tables(18)
Fig. 9. Comparison of adaptive results between the KAIST dataset and the LLVIP dataset
Fig. 10. Comparison of colorization results using different network structures
Fig. 11. Comparison of colorization results using different loss functions
Table 1. Implementation and optimization of CFSA
View tableView in ArticleTable 1. Implementation and optimization of CFSA
def CFSA（Fx）：
F 1 = Conv1×1（F x）F 2 = ChannelAttention（GroupConvolution（F x））+F xQ m = Linear（WindowPartition（F 1））Q n = Linear（WindowPartition（F 2））α，β = LearnableWeights（）
Q = αQ m + βQ nV = Linear（Concat（F 1，F 2））d = Dimension（
Q ）Attention = Softmax（
Q @Q T / sqrt（d））F out = Attention @V Return
F out
Table 2. Implementation and optimization of MFRM
View tableView in ArticleTable 2. Implementation and optimization of MFRM
def MFRM（Fx）：
F x = Conv1×1（F x）F1，F2，F3，F4 = split（F x，num_groups=4）F f1 = Conv1×1（F 1）F f2 = Conv1×1（Conv3×3（F 2））F f3 = Conv1×1（ChannelAttention（F 3））F f4 = Conv1×1（SpatialAttention（F 4））Y 1 = Multiply（F f1，F f2）Y 2 = Multiply（F f1，F f3）Y 3 = Multiply（F f1，F f4）α，β，γ = LearnableWeights（）
Y = αY 1 + βY 2 + γY 3Y = Conv1×1（Y ）Return
Y
Table 3. Performance comparison of various comparison methods on KAIST and LLVIP datasets
View tableView in ArticleTable 3. Performance comparison of various comparison methods on KAIST and LLVIP datasets
Method KAIST Dataset LLVIP Dataset PSNR↑ SSIM↑ FID↓ LPIPS↓ PSNR↑ SSIM↑ FID↓ LPIPS↓ CUT 27.763 0.652 56.18 0.2571 26.549 0.536 63.87 0.3354 IRC 28.691 0.813 54.87 0.1929 27.742 0.615 58.61 0.2912 sRGB-TIR 28.105 0.737 60.45 0.2293 27.138 0.592 58.38 0.3141 MappingFormer 28.978 0.833 51.57 0.1859 28.164 0.637 55.73 0.2781 Proposed 29.195 0.862 50.75 0.1824 28.396 0.651 52.93 0.2521
Table 4. Performance comparison of different network structures for colorization results
View tableView in ArticleTable 4. Performance comparison of different network structures for colorization results
Method Network structure Evaluation index RRAM CFSA MFRM PSNR↑ SSIM↑ FID↓ LPIPS↓ Structure1 √ × × 28.294 0.680 65.30 0.2655 Structure2 × √ × 28.130 0.652 68.46 0.2862 Structure3 × × √ 27.965 0.639 67.84 0.2772 Structure4 × √ √ 28.374 0.736 61.90 0.2171 Structure5 √ × √ 28.529 0.789 56.31 0.2079 Structure6 √ √ × 28.934 0.854 51.32 0.2061 Structure7 √ √ √ 29.195 0.862 50.75 0.1824
Table 5. Performance comparison of different loss functions for colorization results
View tableView in ArticleTable 5. Performance comparison of different loss functions for colorization results
Method Loss function Evaluation index PSNR↑ SSIM↑ FID↓ LPIPS↓ Loss1 √ √ × × 28.523 0.731 63.37 0.2628 Loss2 √ √ √ × 28.984 0.809 56.17 0.2135 Loss3 √ √ × √ 29.013 0.816 53.68 0.2068 Loss4 √ √ √ √ 29.195 0.862 50.75 0.1824
Table 6. Performance comparison of different similarity loss weights for colorization results
View tableView in ArticleTable 6. Performance comparison of different similarity loss weights for colorization results
PSNR↑ SSIM↑ FID↓ LPIPS↓ 1 28.522 0.746 69.53 0.267 1 5 28.948 0.810 56.79 0.216 5 10 29.195 0.862 50.75 0.182 4 15 28.735 0.753 64.98 0.188 4
Tools
Get Citation
Copy Citation Text
Guodong YU, Jianguo ZHU, Chunyang WANG, Jianghai FENG, Xubin FENG, Shi LIU, Pengyu XU, Zhongqi LI, Xiaochen LIU. Combining Generative Adversarial Network and Contrastive Learning for Infrared Image Generation[J]. Acta Photonica Sinica, 2025, 54(5): 0510002
Paper Information
Category:
Received: Oct. 30, 2024
Accepted: Jan. 17, 2025
Published Online: Jun. 18, 2025
The Author Email: Chunyang WANG (fjh879211@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20