Deep Learning-Based Enhancement of Hyperspectral Images Using Simulated Ground Truth

The paper addresses the problem of imaging quality enhancement for the Offner hyperspectrometer using a convolutional neural network. We use a deep convolutional neural network with residual training and PReLU activation, inspired by the super-resolution task for RGB images. In the case of hyperspectral imaging, it is often a problem to find a large enough ground truth dataset for training a neural network from scratch. Transfer learning using the network pretrained for RGB images with some pre- and postprocessing is one of the possible workarounds. In this paper, we propose to simulate the necessary ground truth data using non-imaging spectrometer. The obtained dataset with partially simulated ground truth is then used to train the convolutional neural network directly for hyperspectral image quality enhancement. The proposed training approach also allows to incorporate distortions specific for hyperspectral images into the enhancement procedure. It allows to successfully remove the striping distortions inherent to the Offner scheme of image acquisition. The experimental results of the proposed approach show a significant quality gain.