Hyper Spectral Image Super-Resolution(HSI-SR) has gained significant attention in recent years due to its potential applications.However,the challenge of obtaining high-resolution hyperspectral images is compounded by...Hyper Spectral Image Super-Resolution(HSI-SR) has gained significant attention in recent years due to its potential applications.However,the challenge of obtaining high-resolution hyperspectral images is compounded by limitations in sensor resolution and the high dimensionality of spectral data.Traditional approaches,including interpolation-based methods and sparse representation techniques,often struggle to capture the intricate spectral-spatial dependencies in hyperspectral images.To address these limitations,this study proposes a Hadamard Self-Attention Network(HSAN) for fusing a High-resolution Multispectral Image(Hr-MSI) and a Low-resolution Hyper Spectral Image(Lr-HSI),achieving HSI-SR for obtaining a High-resolution Hyper Spectral Image(Hr-HSI).The core of HSAN is a new Hadamard self-attention mechanism that can be more efficient than traditional dot-product attention because it avoids matrix multiplications and softmax operations.Considering that deep learning-based data fusion typically entails a significant computational and storage burden,this new approach can be integrated with convolutional layers to form an unsupervised lightweight network,which significantly reduces dependence on computational resources.Experimental results across four datasets validate the effectiveness and advantages of HSAN,compared with state-of-the-art approaches.The source code will be available at https://github.com/zxnhkdm/HSAN.展开更多
基金National Natural Science Foundation of China(No. 42571463)Macao Young Scholars Program(No. AM2023033)+1 种基金Shaanxi Province Youth Science and Technology Star Program(No. 2024ZCKJXX-115)Natural Science Foundation of Shaanxi Province(No.2025JC-YBMS-257)。
文摘Hyper Spectral Image Super-Resolution(HSI-SR) has gained significant attention in recent years due to its potential applications.However,the challenge of obtaining high-resolution hyperspectral images is compounded by limitations in sensor resolution and the high dimensionality of spectral data.Traditional approaches,including interpolation-based methods and sparse representation techniques,often struggle to capture the intricate spectral-spatial dependencies in hyperspectral images.To address these limitations,this study proposes a Hadamard Self-Attention Network(HSAN) for fusing a High-resolution Multispectral Image(Hr-MSI) and a Low-resolution Hyper Spectral Image(Lr-HSI),achieving HSI-SR for obtaining a High-resolution Hyper Spectral Image(Hr-HSI).The core of HSAN is a new Hadamard self-attention mechanism that can be more efficient than traditional dot-product attention because it avoids matrix multiplications and softmax operations.Considering that deep learning-based data fusion typically entails a significant computational and storage burden,this new approach can be integrated with convolutional layers to form an unsupervised lightweight network,which significantly reduces dependence on computational resources.Experimental results across four datasets validate the effectiveness and advantages of HSAN,compared with state-of-the-art approaches.The source code will be available at https://github.com/zxnhkdm/HSAN.
