Rethinking 3D-CNN in Hyperspectral Image Super-Resolution

Recently, CNN-based methods for hyperspectral image super-resolution (HSISR) have achieved outstanding performance. Due to the multi-band property of hyperspectral images, 3D convolutions are natural candidates for extracting spatial–spectral correlations. However, pure 3D CNN models are rare to see, since they are generally considered to be too complex, require large amounts of data to train, and run the risk of overfitting on relatively small-scale hyperspectral datasets. In this paper, we question this common notion and propose Full 3D U-Net (F3DUN), a full 3D CNN model combined with the U-Net architecture. By introducing skip connections, the model becomes deeper and utilizes multi-scale features. Extensive experiments show that F3DUN can achieve state-of-the-art performance on HSISR tasks, indicating the effectiveness of the full 3D CNN on HSISR tasks, thanks to the carefully designed architecture. To further explore the properties of the full 3D CNN model, we develop a 3D/2D mixed model, a popular kind of model prior, called Mixed U-Net (MUN) which shares a similar architecture with F3DUN. Through analysis on F3DUN and MUN, we find that 3D convolutions give the model a larger capacity; that is, the full 3D CNN model can obtain better results than the 3D/2D mixed model with the same number of parameters when it is sufficiently trained. Moreover, experimental results show that the full 3D CNN model could achieve competitive results with the 3D/2D mixed model on a small-scale dataset, suggesting that 3D CNN is less sensitive to data scaling than what people used to believe. Extensive experiments on two benchmark datasets, CAVE and Harvard, demonstrate that our proposed F3DUN exceeds state-of-the-art HSISR methods both quantitatively and qualitatively.