Learning To Disentangle Representations For Rain Streak Removal

Rain streak removal is an important task in many real-world vision applications since rain streaks in the air largely threaten the performance of visual analytics. When training conventional models, it is common to remove rain streaks without defining explicitly biased information. Although it is often regarded that learned representation is effective in capturing informativeness, it is still likely to indicate incongruent information to spoil removal while learning biased representation. To handle this issue, we employ an information-theoretic concept to define disentangled representation which is divided into shared and biased characteristics respectively. Our key idea is to remove biased feature representations from a set of co-occurrence features while preserving details using mutual information. We achieve this by proposing 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 five benchmark datasets show the superiority of our new model against state-of-the-art methods.