Neural Network Fragile Watermarking With No Model Performance Degradation

Recently, deep-learning-based methods have made remarkable progress for reconstructing the high-resolution hyperspectral (HR-HS) image through automatically learning the inherent priors from images. These methods are basically implemented in a fully-supervised learning manner with a previously prepared large-scale external dataset captured under controlled conditions, which would greatly restrict the wide applicability to real scenarios. Therefore, this study proposes a novel generalized deep internal learning to solve the HS image super-resolution (HSI SR) problem. Specifically, we aim to train an image-specific CNN model for an under-studying scene using the extracted triplet samples from the down-sampled LR-HS and HR-RGB images and the original LR-HS image as well as the observed HR-RGB and LR-HS images without the corresponding ground-truth for unsupervised learning. To implement the unsupervised learning, we design the degradation blocks to approximate the spatial and spectral degradation operations, and then transform the learned HR-HS target to the LR-HS and HR-RGB estimations for evaluating the network learning states. To verify the effectiveness of our proposed framework, we conduct extensive experiments on two benchmark HS datasets, and demonstrate that the proposed method achieves favorable performance over the state-of-the-art methods.