Year: 2025
Type: article
Abstract: 针对高光谱图像(HSI)与多光谱图像(MSI)融合中光谱信息整合与超分辨率性能的挑战，本文提出了一种基于AutoFusion和循环反馈的高光谱图像融合网络。该网络首先通过光谱分组对低分辨率HSI进行超分辨率重建，利用循环反馈机制迭代更新初始超分辨率图像，结合自适应融合模块(AutoFusion)动态融合多源信息，生成每组的高质量超分辨率特征。其次，通过特征提取网络分别处理融合特征和MSI特征，结合上采样和点卷积操作进一步优化特征表示。最后，将每组特征输入GRU模块生成注意力权重，用于加权调整各组超分辨率结果，重构最终的高分辨率高光谱图像(HR-HSI)。实验在CAVE和Harvard数据集上验证了Net模型，与七种主流融合算法相比，该网络在PSNR、SSIM和SAM等指标上表现优异，证明了其在高光谱图像融合领域的先进性。Aiming at the challenges of spectral information integration and super-resolution performance in the fusion of hyperspectral images (HSI) and multispectral images (MSI), this paper proposes a hyperspectral image fusion network based on AutoFusion and cyclic feedback. This network first performs super-resolution reconstruction of the low-resolution HSI through spectral grouping, iteratively updates the initial super-resolution image using the cyclic feedback mechanism, and dynamically fuses multi-source information in combination with the adaptive fusion module (AutoFusion) to generate high-quality super-resolution features for each group. Secondly, the fused features and MSI features are processed respectively through the feature extraction network, and the feature representation is further optimized by combining upsampling and dot convolution operations. Finally, each group of features is input into the GRU module to generate attention weights, which are used to weighted adjust the super-resolution results of each group and reconstruct the final high-resolution hyperspectral image (HR-HSI). The experiments verified the Net model on the CAVE and Harvard datasets. Compared with seven mainstream fusion algorithms, this network performed outstandingly in indicators such as PSNR, SSIM and SAM, proving its advancement in the field of hyperspectral image fusion.
Source: Modeling and Simulation
Author: 景 梅
Language: English
Subfield: Media Technology
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4411248920