LAUSR

We propose a new deep network architecture for removing a stripe noise from a single meteorological satellite infrared cloud image. In the proposed framework, a residual learning is utilized to directly reduce the mapping range from input to output, which speeds up the training process as well as boosts the destriping performance. Inspired by the wide inference networks, we use wider CNNs with more convolutions in the first part of the proposed network, which is helpful for learning the similar pixel-distribution features from noisy images. To further improve the performance, we propose a local-global combination structure model, which combines the representations of different layers for recovering the rich details of infrared cloud images. Moreover, we extend our method to remove rain streaks from single images, which provides a new idea for rain-removal task. In addition, we provide a new meteorological satellite infrared cloud image dataset for training and validating the proposed network. Final extensive experiments demonstrate that the proposed method can achieve both comparable restoration quality and computational efficiency with several state-of-the-art approaches.