图像去雾中深度学习数据增强方法

A Deep Learning Data Enhancement Method in Image Haze Removal

图像去雾是图像处理领域中非常重要的问题。深度学习可以有效提高图像清晰度,但训练过程中由于缺少相对应的真实雾匹配数据对,多采用合成雾作为数据集。现有合成雾多依赖于深度信息、大气散射系数等参数,针对由此作为数据集训练容易造成颜色失真和去雾不彻底的问题,提出基于循环生成对抗网络(CycleGAN)合成雾方法。通过该网络进行不匹配数据对训练学习有雾图像的特征,然后赋予清晰图片真实雾特征并与其自身构成匹配数据对,最后再用此类数据集进行去雾训练。结果表明,这些数据集可以有效解决颜色失真和去雾不彻底等问题。

Image haze removal is an important issue in the field of image processing.Deep learning can ffctively improve image clarity,but due to the lack of corresponding real haze matching data pairs in the training process,synthetie haze is usually used as the dataset.The existing synthetie haze mostly depends on such parameters as depth information and atmospheric scattering cofficient.To solve the problems of color distortion and incomplete haze removal caused by training on such a dataset,a synthetie haze method based on Cycle Generative Adversarial Network(CycleGAN)is proposed.Through the network,mismatched data pair training is conducted to learn the features of haze images,then real haze features are added to clear pictures,with which matching data pairs are formed,and finally such datasets are used for dehazing training.The results show that these datasets can ffectively solve the problems of color distortion and incomplete haze removal.