基于深度残差卷积神经网络的高光谱图像超分辨方法

卷积神经网络由于其强大的非线性表达能力在自然图像的处理问题中已经获得了非常大的成功。传统的稀疏表示方法利用精确配准的高分辨率多光谱图像，从而限制了实际应用。针对传统方法的不足，本文提出了一种基于深度残差卷积神经网络的单高光谱图像超分辨率方法，无需对应的多光谱图像。我们构建深度残差卷积神经网络挖掘低分辨率遥感图像和高分辨率遥感图像之间的非线性关系。构建的深度学习网络串联多个残差块，并去除一些不必要的模块，如批标准化层，每个残差块只包含两个卷积层，这样在保证模型效果的同时又加快模型的效率。此外，因为遥感图像训练数据缺乏，我们充分挖掘自然图像和高光谱图像之间的相似性，利用自然图像样本训练卷积神经网络，进一步利用迁移学习将训练好的网络模型引入到高分辨率遥感图像超分辨问题上，解决了训练样本缺乏问题。最后，基于实际的遥感数据超分辨实验结果表明，本文所提出的方法具有良好的性能，能得到较好的超分辨效果。

Hyperspectral image super-resolution method based on deep residual convolutional neural network

Convolutional neural networks have achieved great success in the processing of natural images due to their pow-erful nonlinear expression capabilities. Traditional sparse representation methods use high-resolution mul-ti-spectral images with precise registration, which limits their practical applications. Aiming at the shortcomings of traditional methods, this paper proposes a single hyperspectral image super-resolution method based on deep re-sidual convolutional neural network without the need for corresponding multispectral images. We construct a deep residual convolutional neural network to mine the nonlinear relationship between low-resolution remote sensing images and high-resolution remote sensing images. The constructed deep learning network concatenates multiple residual blocks and removes some unnecessary modules, such as batch normalization layers, each residual block contains only two convolutional layers, so as to ensure the model effect and speed up the model efficiency. In addition, due to the lack of training data for remote sensing images, we fully mine the similarities between natural images and hyperspectral images, use natural image samples to train convolutional neural networks, and further use transfer learning to introduce trained network models to high-resolution remote sensing image su-per-resolution problem, the problem of lack of training samples is solved. Finally, the results of super-resolution experiments based on actual remote sensing data show that the proposed method has good performance and can obtain good super-resolution results.