Prior-combined dehazing network based on mutual learning

Single-image dehazing is an important problem for high-level computer vision tasks since the existence of haze severely degrades the recognition ability of computers. Most recent works tend to combine prior-based dehazing method with a convolutional neural network to improve the dehazing effect in real scenes. However, these methods do not tackle with the color shifts caused by prior-based methods effectively. In this paper, we propose a prior-combined dehazing network based on mutual learning. Specifically, we build two sub-networks to achieve dehazing by both supervised and unsupervised ways. The supervised sub-network is optimized by ground truth, which provides color fidelity but may acquire under-dehazed images when applied to real scenes. The unsupervised sub-network is optimized by the dehazed images of dark channel prior, which improves the generalization ability but introduces some color shifts or artifacts. Since the dehazing of these two sub-networks shows complementary advantages, a mutual learning mechanism is built for the joint optimization. And we propose a feature fusion module based on the perceptual differences to acquire the final results. The experimental results demonstrate that our method surpasses previous state-of-the-arts on both synthetic and real-world datasets.