Incremental Image De-raining via Associative Memory

While deep learning models have achieved the state-of-the-art performance on single-image rain removal, most methods only consider fixed mapping rules single synthetic dataset for lifetime. This limits real-life application as iterative optimization may change and training samples. However, when learn a sequence of datasets in multiple incremental steps, they are susceptible to catastrophic forgetting that adapts new episodes while failing preserve previously acquired rules. In this paper, we argue importance sample diversity optimization, propose novel memory management method, Associative Memory, achieve image de-raining. It bridges connections between current past feature reconstruction by sampling domain mappings guides trace pathway back historical environment without storing extra data. Experiments demonstrate our method can better than existing approaches both inhomogeneous within spectrum highly compact systems.