【摘要】 为使雾天拍摄照片清晰,改善合成数据集泛化能力不足的缺点,提出了一种基于半监督学习的图像去雾算法。算法包含了监督训练和无监督训练两个部分,分别使用人工合成数据集和真实有雾图像交替训练网络,在网络中引入了平滑扩张卷积来提高感受野并消除网格伪影。无监督训练部分将生成对抗网络作为基础,采用马尔可夫判别器以提高网络对细节的恢复能力。实验结果表明,所提算法在去雾程度、纹理清晰度等方面都有所提升,提高了算法对真实图像的泛化能力。更多还原

【Abstract】 In order to make the photos taken in foggy days clear and improve the shortcomings of insufficient generalization ability of synthetic datasets,an image dehazing algorithm based on semi-supervised learning is proposed.The algorithm includes supervised training and unsupervised training,using artificial synthetic data sets and real haze images to alternately train the network.And the smooth dilated convolution is introduced in the network to improve the receptive field and eliminate grid artifacts.The unsupervised training part is based on the generation of adversarial networks,and Markov discriminator is used to improve the network’s ability to recover details.The experimental results show that the proposed algorithm has improved the degree of haze removal and texture definition,which improves the generalization ability of the algorithm to real images.更多还原