An Efficient Residual Learning Neural Network for Hyperspectral Image Superresolution

摘要

Deep learning, especially a discriminative model for image reconstruction, has shown great potential for single image superresolution (SR) of hyperspectral images (HSI). For HSI SR task, it is crucial to predicting each pixel according to the surrounding context, exploiting both spatial and spectral correlation information simultaneously. In this paper, an efficient three-dimensional (3-D) HSI SR convolution neural network (CNN) based on residual learning is proposed. The network builds convolutional layers in low-resolution (LR) space and extracts the features along both spatial and spectral dimensions using 3-D dilated kernel. Then, 3-D deconvolution is employed at the last layer, which enlarges the image to the desired size. By employing multibranch and multiscale fusion in the architecture, the network can learn a better and more complex LR to high-resolution mapping. The overall network combines the global with local residual learning to reduce training difficulty and improve the performance. The design philosophy of our model is to find the best tradeoff between performance and computational cost. We train the network in an end-to-end fashion, and the experimental results of the quantitative and qualitative evaluation show that our proposed method yields satisfactory SR performance.