From depth-aware haze generation to real-world haze removal

For deep learning-based single image dehazing works, their performances seriously depend on the designed models and training dataset. Existing state-of-the-art methods focus on the design of novel dehazing models or the improvement of training strategies to obtain better dehazing results. In this work, instead of designing a new deep dehazing model, we attempt to further improve the dehazing performance from the perspective of enriching training datasets by exploring an intuitive yet efficient way to synthesize photo-realistic hazy images. It is well known that for a natural hazy image, its perceived haze density increases with scene depth. Motivated by this, we develop a depth-aware haze generation network, namely HazeGAN, by incorporating the Generative Adversarial Network (GAN), depth estimation network, and physical atmospheric scattering to progressively synthesize hazy images. Specifically, a separate depth estimation network is embedded to obtain multi-scale depth features, which are exploited by the atmospheric scattering model to generate multi-scale hazy features. The hazy features are fused into the GAN generator to output synthetic hazy images with depth-aware haze effects. Extensive experimental results demonstrate that the proposed HazeGAN can generate diverse training pairs of depth-aware hazy images and clear images, which effectively enrich the existing benchmark datasets, and improve the generalization capabilities of existing deep image dehazing models.