From synthetic to natural — single natural image dehazing deep networks using synthetic dataset domain randomization

Image dehazing methods aim to solve the problem of poor visibility in images due to haze. Techniques proposed for image dehazing in literature focus on image priors, haze lines or data driven statistical models. Variations of the classical methods relying on prior model or haze line model use no-reference image quality metrics to prove their dehazing performance. Recently developed deep learning models rely on huge amounts of hazy, haze-free pairs for training, and uses PSNR and SSIM like image reconstruction metrics to show their performance. These methods perform poorly on no-reference image quality assessments and also dehazes poorly at the depths of the image. These methods though can be optimized for memory usage and are faster. This work presents a deep learning model (Feature Fusion Attention Network) trained on a domain randomized synthetic dataset generated in simulation. The proposed model achieves the highest scores on blind image assessments through the gradient rationing technique for a deep learning-based approach by a significant margin. The images were evaluated on full-reference metrics as well and obtained favorable results. This approach also yields one of the highest edge sharpness obtained after dehazing. The training procedure adopted to obtain significant gains on real-world dehazing, without using any real-world data is also detailed in this paper.