Single Image Defogging Method based on Deep Learning

Abstract

Abstract. Single image defogging is a challenging ill-posed problem. Current image defogging methods usually get defogging solutions based on various priors or assumption, which is hardly satisfied in practice. In this paper, a single image defogging method based on deep learning is proposed, in which the priors and assumption do not hold. Firstly, the prediction of transmission map is progressively refined by using three scales convolutional neural networks. Secondly, the fog-free image can be recovered by the atmospheric scattering model after transmission map is got. Experiment results show that the proposed deep learning based defogging algorithm achieves superior performance over state-of-art algorithms on both natural foggy images and synthetic images.

Image Defogging Framework Using Segmentation and the Dark Channel Prior

“…Therefore, it cannot be applied on real world images without prior knowledge. In [29], three scales convolutional neural network to predict transmission map is used. Then, fog free images are recovered using the atmospheric scattering model.…”