Enhancing Rainy Image via Invertible Networks

Rainy-day images often suffer from significant degradation in visual quality due to weather-related interference. The task of removing rain streaks from rainy images is referred to as image deraining, which is highly challenging due to the weak distinction between rain streaks and background images. This often leads to the loss of image details during the rain removal process. To address this issue, this paper innovatively proposes a learning-based single-image deraining method (Inv-DeRain) using reversible neural networks, which can accurately eliminate rain streaks while effectively preserving the fine texture details of the image. Specifically, we formulate image deraining as an image decomposition problem and propose an invertible neural network (INN) with an asymmetric interactive architecture, which separately predicts the rain streak layer and the background image. This INN progressively decouples the background and rain layers while preserving fine image details. Based on the reversible characteristics of Inv-DeRain, we designed a two-way manner loss function to facilitate the network's learning of detailed image deraining knowledge and mitigate the impact of rain streaks resembling the background. Our extensive experiments demonstrate that the proposed method achieves performance on par with state-of-the-art image deraining techniques on benchmark datasets, while exhibiting superior quantitative results in terms of generalization capability.