A Pixel-to-Pixel Convolutional Neural Network for Single Image Dehazing

Estimating transmission maps is the key to single image dehazing. Recently, Convolutional Neural Networks based methods (CNNs), which aim to minimize the difference between the predictions and the transmission maps, have achieved promising dehazing results and outperformed traditional feature-based algorithms. However, two transmission maps with the same estimation error can produce quite different dehazing results. Therefore, these models are incapable to directly affect the quality of the restorations. To address this issue, we propose a pixel-to-pixel dehazing convolutional neural network in this paper, which learns a map from the hazy images to the haze-free screens. Specifically, we intuitively maximize the visual similarity between the predicted images and the ground truth with some visual-relevant loss functions, e.g., the mean square error and the gradient difference loss. Experiments on synthetic dataset and real images demonstrate that our method is effective and outperforms the state-of-the-art dehazing methods.