Bidirectional deep residual learning for haze removal

초록

Recently, low-level vision problems has been addressed using residual learning that can learn a discrepancy between hazy and haze-free images. Following this approach, in this paper, we present a new dehazing method based on the proposed bidirectional residual learning. Our method is implemented by generative adversarial networks (GANs), consisting of two components, namely, haze-removal and haze-reconstruction passes. The method alternates between removal and reconstruction of hazy regions using the residual to produce more accurate haze-free images. For efficient training, we adopt a feature fusion strategy based on extended tree-structures to include more spatial information and apply spectral normalization techniques to GANs. The effectiveness of our method is empirically demonstrated by quantitative and qualitative experiments, indicating that our method outperforms recent state-of-the-art dehazing algorithms. In particular, our approach can be easily used to solve other low-level vision problems such as deraining. © 2019 IEEE Computer Society. All rights reserved.