Hyperspectral Images Reconstruction based on the 3D Fast Super-Resolution Convolutional Neural Networks

Abstract

Hyperspectral images are widely used with the development of remote sensing technology, but its low spatial resolution greatly limits the application of hyperspectral images. Therefore, we proposed a hyperspectral image super-resolution reconstruction algorithm based on 3d FSRCNN framework model, which preprocessed the hyperspectral image data set successively, extracted features, reduced dimensions, nonlinear mapping, expanded dimensions, and deconvolution to finally realize image super-resolution reconstruction. In this algorithm, three-dimensional convolution is used to convolve both spatial dimension and spectral dimension to capture spatial spectrum characteristics. The whole convolutional network consists of 6 layers, one input layer, four convolutional layers, and one deconvolution layer. For the first five layers of convolution, PReLU was used as the activation function, which effectively prevented the phenomenon of nerve necrosis and improved the model fitting without increasing the computational cost and overfitting risk. Experimental results show that the proposed algorithm can reconstruct high spatial resolution images with less computation and reduce spectral distortion effectively.