高光谱成像
计算机科学
人工智能
计算机视觉
图像(数学)
对偶(语法数字)
图像处理
遥感
模式识别(心理学)
图像分割
迭代重建
人工神经网络
特征提取
图像分辨率
合成孔径雷达
算法设计
数学形态学
作者
Kewen Huang,Yiping Xu,Dong Yi,Ke Li,Hu Wang,Yan Tian
标识
DOI:10.1109/tgrs.2026.3678075
摘要
To address the problem that current hyperspectral image super-resolution methods struggle to effectively extract spatial-spectral information, resulting in degraded performance, this paper proposes a dual-branch network for hyperspectral images super-resolution (DBNSR). In the proposed framework, one branch transforms hyperspectral image super-resolution into abundance map super-resolution through unmixing, leveraging endmembers to represent spectral information and thereby enhancing spectral feature extraction. Simultaneously, the other branch performs super-resolution directly on the spatial domain of the original hyperspectral image to effectively extract spatial features. Furthermore, a dedicated high-frequency feature extraction module is proposed to capture fine-grained details via spatial-spectral residual blocks, thereby enhancing superresolution performance. Finally, a spatial-spectral feature fusion module is introduced to integrate the outputs of the dual branches, enabling effective fusion of spatial and spectral information. Experimental results on three datasets demonstrate that the proposed method outperforms current state-of-the-art methods both metrically and visually. The code is available at https://github.com/fomanwong/DBNSR.
科研通智能强力驱动
Strongly Powered by AbleSci AI