基于多先验约束和一致性正则的半监督图像去雾算法

针对合成雾霾图像训练的去雾模型在真实场景中去雾效果不佳、对高层视觉任务性能提升不明显等问题，该文提出一种基于多先验约束和一致性正则的半监督图像去雾算法。该方法采用编码器-解码器网络结构，同时在合成雾霾图像与真实雾霾图像上学习去雾映射，并利用多种统计先验去雾结果作为真实雾霾图像参考真值进行半监督学习，同时通过多张真实雾霾图像的随机混合进行一致性正则约束，以消除多种先验去雾结果差异以及噪声干扰，提高图像去雾结果的视觉质量。实验对比结果表明，所提算法可比现有方法获得更好的真实场景去雾结果，并且能够显著提升高层视觉任务性能。

Semi-supervised Image Dehazing Algorithm Based on Multi-prior Constraint and Consistency Regularization

Previous dehazing models trained on synthetic hazy images can not generalize well on real hazy scenes and improve the performance of high-level vision tasks significantly. To resolve this issue, a semi-supervised image dehazing based on multi-priors constrain and output consistency regularization is proposed. The algorithm adopts the encoder and decoder network to train on the synthetic and real hazy images by sharing the parameters. Multi prior-based dehazed images are adopted as pseudo labels to constrain the real scene hazy images. Furthermore, to reduce the divergence of different prior-based methods, the dehazing results of the random mix-up real hazy images are regularized to be consistent with the corresponding mix-up of the prior-based dehazed images. Finally, the experiment results demonstrate the performance of the proposed algorithm compared with the state-of-the-art methods.