Recursive Deep Residual Learning for Single Image Dehazing

Abstract

There have been a flurry of works on deep learning based image dehazing in recent years. However, most of them have only used deep neural networks to estimate the transmission map (or its variant); while the module of generating dehazed image is still model-based. Inspired by the analogy between image dehazing and image denoising, we propose to reformulate dehazing as a problem of learning structural residue (instead of white Gaussian noise) and remove haze from a single image by a deep residue learning (DRL) network. Such novel reformulation enables us to directly estimate a nonlinear mapping from input hazy images to output dehazed ones (i.e., bypassing the unnecessary step of transmission map estimation). The dehazing-denoising analogy also motivates us to leverage the strategy of iterative regularization from denoising to dehazing - i.e., we propose to recursively feed the dehazed image back to the input of DRL network. Such recursive extension can be interpreted as a nonlinear optimization of DRL whose convergence can be rigorously analyzed using fixed-point theory. We have conducted extensive experimental studies on both synthetic and real-world hazy image data. Our experimental results have verified the effectiveness of the proposed recursive DRL approach and shown that our technique outperforms other competing methods in terms of both subjective and objective visual qualities of dehazed images.