关键词: 图像去雾, 深度学习, 卷积神经网络, 残差学习, 端到端系统

Abstract: Considering the disadvantage of traditional dehazing algorithm, a single image dehazing algorithm based on haze layers learning is proposed. According to the atmospheric scattering model, the end-to-end network is designed which directly learn the mapping between the haze images and their corresponding haze layers. The network takes the haze image as the input. Then the recovered haze-free image can be gotten by removing the residual image from the hazy image. Residual learning allows the network to estimate the initial haze layer with relatively high learning rates, which can reduce computational complexity and speed up the convergence process. Otherwise, we use guided filter to refine images avoiding halos and block artifacts, which make the recovered image more similar to the real scene. Finally, the experimental results are analyzed and contrasted carefully. In this paper, the effect on fog images with different fog density is tested, and many comparisons are listed with other classical algorithms. Experiments demonstrate that the proposed algorithm has better results than state-of-the-art methods on both synthetic and real-world images qualitatively and quantitatively.

Key words: image dehaze, deep learning, convolutional neural network, residual learning, end-to-end system