计算机科学
人工智能
计算机视觉
高光谱成像
模式识别(心理学)
图像融合
人工神经网络
融合
图像(数学)
图像处理
传感器融合
特征提取
图像分割
信号处理
目标检测
图像配准
作者
Qiya Song,Zhipeng Xie,Shutao Li,Renwei Dian,Chengxin Xie,Jingui Huang
标识
DOI:10.1109/tmm.2026.3744789
摘要
Hyperspectral and multispectral image fusion aims to reconstruct high-resolution hyperspectral images that simultaneously preserve fine spatial details and high spectral fidelity. However, existing methods struggle to jointly ensure spatial detail reconstruction and spectral consistency: 1) The shallow modeling of directional geometric priors impedes the effective representation of high-frequency spatial structures, leading to structural deviation; 2) The cross-modal fusion process tends to focus on spatial information while sacrificing spectral information, leading to an imbalance in spatial-spectral modeling and making it difficult to sustain spectral consistency. To overcome these limitations, we propose a Hierarchical Gradient Aware Network (HGANet), which aggregates information across different spatial resolutions via a multi-scale fusion strategy, explicitly integrates multi-level gradient priors with adaptive gated attention for robust multi-scale spatial-spectral fusion. Specifically, a Hierarchical Gradient Interaction Module (HGIM) is designed to extract and refine directional gradient features from both hyperspectral and multispectral branches, injecting hierarchical gradient residuals into the fusion backbone to mitigate cumulative structural deviation. Moreover, an Adaptive Cross-modal Gated Fusion Module (ACGFM) is designed to integrate hybrid channel-spatial attention with a multi-layer perceptron (MLP)-based gating mechanism, where dynamic weighting is performed on modal spatial-spectral features to balance the spatial-spectral complementary information throughout the fusion process and thereby enhance spectral consistency. Extensive experiments on both simulated and real satellite datasets demonstrate that HGANet outperforms state-of-the-art methods in terms of quantitative accuracy and perceptual quality, achieving superior spectral fidelity and spatial sharpness.
科研通智能强力驱动
Strongly Powered by AbleSci AI