摘要
为解决现有红外与可见光图像融合目标不够显著,轮廓纹理细节不够清晰等问题,提出了一种基于交互注意力的红外和可见光图像融合网络。该方法通过双流特征提取分支提取源图像的多尺度特征,然后经过交互融合网络获得注意力图,以便从红外与可见光图像中自适应地选择特征进行融合,最后通过图像重建模块生成高质量的融合图像。在MSRS数据集和TNO数据集的实验结果表明,所提方法在主观视觉描述和客观指标评价方面均表现出了较好的性能,融合结果包含更清晰的细节信息和更明显的目标。
Considering insufficient remarkability and unclear contour texture of existing infrared and visible images,an interactive attention-based infrared and visible image fusion network was proposed.In which,multi-scale features of the source image through a dual-stream feature extraction branch was extracted and an attention map was obtained through the interactive fusion network to adaptively select features from the IR and visible images for fusion and finally generates a high-quality fused image through the image reconstruction module.Experiments on both the MSRS dataset and the TNO dataset show that,the algorithm proposed exhibits better performance in both subjective visual description and objective index evaluation,and the image fusion results contain clearer detail information and more obvious targets.
作者
山子岐
邹华宇
李凡
刁悦钦
SHAN Zi-qi;ZOU Hua-yu;LI Fan;DIAO Yue-qin(Faculty of Information Engineering and Automation,Kunming University of Science and Technology)
出处
《化工自动化及仪表》
CAS
2024年第3期523-527,534,共6页
Control and Instruments in Chemical Industry
关键词
图像融合
深度学习
交互注意力
密集残差连接
image fusion
deep learning
interactive attention
dense residual connectivity