Dark Channel Prior Guided Conditional Generative Adversarial Network for Single Image Dehazing

Abstract

Single image dehazing is a challenging while important problem as the existence of haze hinders most high-level computer vision tasks. Previous methods solve this problem using various hand-crafted priors or by CNN learning on synthetic data sets. In practice, many CNN based methods estimate the transmission maps and atmospheric lights without considering the pre-defined priors, and always need huge data to train the model. In this work, we propose Dark Channel Prior Guided Conditional Generative Adversarial Network, an end-to-end model that generates realistic haze-free images using hazy image input and dehaze image based on dark channel prior. A Siamese like encoder is proposed to extracted the feature, and multi-scale features are enhanced by feature aggregation block for decoding. Our algorithm can efficiently combine the prior-based and CNN based image dehazing method. Experimental results on synthetic datasets and real-world images demonstrate our model can generate more perceptually appealing dehazing results, and provide superior performance compared with the state-of-the-art methods.