Learning to See in the Rain via Disentangled Representation

摘要

Image de-raining is an important task in many robot vision applications since rain effects and hazy air largely threaten the performance of visual analytics. While deep neural de-raining models have greatly boosted performance by learning rich representations of rainy input data, they are still likely to indicate incongruent information to spoil de-raining. We find that the process of extracting the feature representations in the rainy image space is surprisingly important for perceptual de-raining performance. We further probe the limitations of the universal restoration learning method and formulate explicitly relevant and irrelevant information onto feature representations. To handle this issue, we employ an information-theoretic concept to define disentangled representation which is divided into shared and excluded characteristics. Our key idea is to remove excluded feature representations from a set of co-occurrence features while preserving details using mutual information. To achieve this, we propose a novel stage-wise training strategy that captures a more discriminative and pure factor that preserves details. Specifically, we utilize an adversarial objective that explicitly defines each representation to enforce disentanglement. Extensive computational experiments on six benchmark datasets show the superiority of our new model against state-of-the-art methods.