基于多尺度密集特征融合的生成式对抗除雾网络

在真实雾天场景下，针对除雾网络无法去除远处雾气、天空区域容易出现噪声的问题，提出了一种基于多尺度密集特征融合的生成式对抗除雾网络，并采用制作的合成雾天数据集进行对抗训练.首先，对除雾网络进行设计，构建了网络模型；其次，从合成晴朗天气图像中利用深度标签生成逼真的雾天数据集，以适用于真实雾天除雾领域；最后，在真实雾天数据集上测试，选取近几年具有代表性的6种基于深度学习的除雾网络进行主观视觉效果，并借助除雾领域常用的无参考图像质量评价指标进行客观分析.研究结果表明:提出的除雾网络在真实场景下的除雾效果较其他网络有显著提升，主观视觉效果明显优于对比的除雾网络，在无参评价指标上综合表现优于其他除雾网络.

Generative Adversarial Network Based on Multi-scale Dense Feature Fusion for Image Dehazing

In view of the poor dehazed effect of the existing dehazing networks in real hazy image and the obvious noise in the sky area of the image， a generative adversarial network based on multi-scale dense feature fusion for image dehazing is proposed. The dehazing network uses the produced synthetic foggy data set for adversarial training. Firstly， the dehazing network is designed and the network model is constructed; secondly， a realistic foggy data set is directly generated from the synthetic sunny weather image using deep tags to be suitable for the dehazed field; finally， the network is tested on the real foggy day data set and selects six representative deep learning dehazing networks in recent years for comparison， and non-reference image quality evaluation indicators are used for objective analysis. The research results show that the effect of the proposed dehazing network in real scenes is significantly improved compared to the other networks. The subjective visual effect is significantly better， and the comprehensive performance is better than the other networks in non-reference image quality evaluation indicators.