Rectificatory Semantic Information Supplement Network(RSIS-net) For Dynamic Scene Deblurring

In this paper, we construct a Generative Adversarial Networks(GANs) to handle the problem of dynamic deblurring. The generator is based on Feature Pyramid Network (FPN-net), and channel-level correction is used in the encoding process of feature extraction to correct the deep features of blurred image. Since FPN-net exists semantic degradation Phenomenon, we repeatedly supplement the corrected semantic information in the decoding path of the generator. We named our algorithm as Rectificatory Semantic Information Supplement Network (RSIS-net), which contains the Semantic Information Rectification mechanism (SIR), the Semantic Information Supplement mechanism (SIS), and a more suitable low-level perception loss function(Ploss) of a deblurring task. Finally, the experiment results show that our RSIS-net outperforms the many state-of-the-art methods. In addition, the processing time of our lightweight network on a single frame is less than 0.5 seconds, which means that the algorithm supports real-time deblurring.