HCT: a hybrid CNN and transformer network for hyperspectral image super-resolution

Abstract

Recently, convolutional neural network (CNN) and transformer based on hyperspectral image super-resolution methods have achieved superior performance. Nevertheless, this is still an important problem how to effectively extract local and global features and improve spectral representation of hyperspectral image. In this paper, we propose a hybrid CNN and transformer network (HCT) for hyperspectral image super-resolution, which consists of a transformer module with local–global spatial attention mechanism (LSMSAformer) and a convolution module with 3D convolution (3DDWTC) to process high and low frequency information, respectively. Specifically, in the transformer branch, the introduced attention mechanism module (LSMSA) is used to extract local–global spatial features at different scales. In the convolution branch, 3DDWTC is proposed to learn local spatial information and preserve the spectral features, which can enhance the representation of the network. Extensive experimental results show that the proposed method can obtain better results than some state-of-the-art hyperspectral image super-resolution methods.