基于双鉴别器生成对抗架构的高光谱图像融合算法  

Hyperspectral Image Fusion Algorithm Based on Dual Discriminator Generative Adversarial Architecture

作　　者：王世盛 沈慧芳 方立 WANG Shisheng;SHEN Huifang;FANG Li(College of Mechanical and Electrical Engineering,Fujian Agriculture and Forestry University,Fuzhou350100,CHN;Quanzhou Institute of Equipment Manufacturing,Haixi Institutes,Chinese Academy of Sciences,Quanzhou Fujian36221,CHN)

机构地区：[1]福建农林大学机电工程学院,福州350100 [2]中国科学院海西研究院泉州装备制造研究中心,福建泉州362216

出　　处：《光电子技术》2025年第1期18-27,51,共11页Optoelectronic Technology

基　　金：国家自然科学基金青年科学基金项目(42101359);福建省高层次人才创新创业项目(2020C003R)。

摘　　要：提出一种基于双鉴别器生成对抗架构的高光谱图像融合算法。生成器网络将编‑解码器架构与空谱联合注意力结合,以增强捕获空间和频谱特征的能力。通过让生成器分别与空间鉴别器和光谱鉴别器建立对抗博弈,从而在保持光谱精度的同时提高空间分辨率。此外,设计了一种通道交叉融合方法用于增强网络中不同层次特征图之间的光谱融合能力。基于CAVE、Pavia University、Cuprite Mine数据集,将所提方法与其他4种先进算法进行对比。实验结果表明,在8倍缩放因子下,所提方法在三个数据集上的峰值信噪比分别达到46.394,43.166和10.406,在视觉效果和客观定量评价上均优于其他算法。In the case of a significant disparities in spatial resolution between hyperspectral and multispectral images,the fusion results often suffered from pronounced distortion issues.A hyperspectral image fusion algorithm was proposed based on a dual-discriminator generative adversarial architecture.The generator network combined an encoder-decoder architecture with spatial-spectral joint attention,furtherly augmenting its ability to capture both spatial and spectral features.By engaging the generator in adversarial games with both spatial and spectral discriminators,the algorithm ensured an enhancement in spatial resolution while preserving spectral accuracy.To improve spectral fusion across different levels of feature maps,a channel cross-fusion method was designed.Based on the CAVE,Pavia University,and Cuprite Mine datasets,the proposed method was compared with four other advanced algorithms.The experimental results showed that at scaling factor 8,the proposed method could achieve peak signal-to-noise ratios of 46.394,43.166,and 10.406 on the three datasets,outperforming other algorithms in both visual quality and objective quantitative evaluation.

关 键 词：生成对抗网络 高光谱图像融合 通道交叉融合 空谱联合注意力 

分 类 号：TN919.8[电子电信—通信与信息系统]