ZRD-Net: A Zero-shot Low-Light Image Dehazing Network

In this paper, we investigate two less covered challenging problems in single image dehazing neural networks, namely how to remove fog from a given image in an unsupervised and zero-shot manner. To this end, we propose a new method based on the idea of layer disentanglement, which considers the blurred image as the entanglement of several “simpler” layers, namely, the unblurred image layer, the transmission layer, the atmospheric light layer, the reflected light layer, and the noise light layer. The main advantages of the proposed ZRD-Net are two-fold. Firstly, it is an unsupervised method that does not use any clean images including fuzzy-clean pairs as ground truth. Second, ZRD is a “zero-shot” method that uses only a single observed blurred image for learning and inference. In other words, it does not follow the traditional paradigm of training deep models on large-scale datasets. These two advantages enable our method to avoid the labor-intensive data collection and domain shift problems of using synthetic blurred images to deal with real images. Extensive comparisons show that our approach performs well in terms of qualitative and quantitative evaluation compared to multiple methods.