Multi Scale Mixed Convolutional Network for Hyperspectral Image Super Resolution

Thanks to the tremendous success of deep learning in natural image super resolution, deep learning-based hyperspectral image super-resolution (HSI SR) methods have been widely researched. In recent years, many scholars attempt to design deeper network structures to reconstruct higher quality HSIs. However, these approaches not only fails to fully exploit the feature information of the images but also introduces a large number of model parameters. Therefore, this paper proposes a multi scale mixed convolutional network for HSI SR processing. It fully extracts image feature information by integrating different scale features of HSI and utilizes 2D/3D mixed convolutions to simultaneously focus on the spatial and spectral information of HSI. To alleviate the computational burden, separable convolutions are used instead of conventional 2D/3D convolutions. Extensive experiments conducted on two HSI datasets demonstrate that this method outperforms existing methods with superior performance.