Unpaired image super-resolution using a lightweight invertible neural network

Unpaired image super-resolution (SR) has recently attracted considerable attention in the unsupervised SR community. In contrast to supervised SR, existing unpaired SR methods inevitably resort to the generative adversarial network (GAN) to explore data distribution on the given HR and unpaired LR dataset. Nevertheless, predominant strategies often strive for sophisticated network structures or training pipelines, making them intractable to apply in real-world scenarios. In this work, a lightweight invertible neural network (INN) is proposed for unpaired SR to alleviate this limitation. Specifically, we regard image degradation and SR as a pair of mutually-inverse tasks and replace the two generators in one-stage GAN with INN. Due to the information lossless nature of INN, it is impossible to generate noise in vain during image degradation. We thus design a simple noise injection network to induce realistic noise, thereby simulating real LR images. To further maintain the stability and realism of the noise, we propose to extract the noise prior from the real-world LR image. With extracted noise prior as input, our noise injection network can narrow the gap between the generated noise and the real one, thereby encouraging the degraded images to match the real-world LR domain. Extensive experiments demonstrate that our method achieves comparable performance with other SOTA methods in quantitative and qualitative evaluations while enjoying faster speed and much smaller parameters.