Uncertainty-Driven Dehazing Network

Deep learning has made remarkable achievements for single image haze removal. However, existing deep dehazing models only give deterministic results without discussing the uncertainty of them. There exist two types in models: aleatoric that comes from noise inherent observations and epistemic accounts model. In this paper, we propose a novel uncertainty-driven network (UDN) improves by exploiting relationship between uncertain confident representations. We first introduce an Uncertainty Estimation Block (UEB) to predict together. Then, Uncertainty-aware Feature Modulation (UFM) block adaptively enhance learned features. UFM predicts convolution kernel channel-wise modulation cofficients conitioned on weighted representation. Moreover, develop self-distillation loss improve representation transferring knowledge one. Extensive experimental synthetic datasets real-world images show UDN achieves significant quantitative qualitative improvements, outperforming state-of-the-arts.