多模型融合的对抗网络图像去雾研究

Research on Defogging of Adversarial Network Images Based on Multi-Model Fusion

摘要: 为了应对单一网络模型在雾天场景下去雾效果不理想的，去雾后图像颜色失真等问题，提出一种基于循环生成对抗网络(CycleGAN)的多去雾模型融合的去雾算法。首先，在整体去雾模型上，针对不同雾气浓度，将传统的单一模型改为多去雾模型，分别将不同浓度的雾气图像传入对应的去雾网络，接着在生成器模块中引入CBAM注意力机制，帮助模型可以更好的考虑通道上和空间上的影响，赋予重要特征以更多的权重，最后针对生成的无雾图像颜色失真问题，提出一种基于Lab色彩空间的新的色彩损失函数。实验结果表明，与传统的CycleGAN模型相比，本文算法在公共合成数据集和真实世界的图像上取得了更好的性能和更好的视觉效果。

Abstract: In order to deal with the problems such as the unsatisfactory effect of single network model in foggy scenes and the image color distortion after fog removal, a new algorithm based on the fusion of mul-tiple fog removal models based on CycleGAN was proposed. First of all, in terms of the overall de-fogging model, the traditional single model is changed into a multi-de-fogging model according to different fog concentrations, and the fog images of different concentrations are respectively intro-duced into the corresponding de-fogging network. Then, CBAM attention mechanism is introduced into the generator module to help the model better consider the influence on channel and space, and give more weight to important features. Finally, a new color loss function based on Lab color space is proposed to solve the color distortion problem of the generated fog free image. The exper-imental results show that compared with the traditional CycleGAN model, the proposed algorithm achieves better performance and better visual effect on public synthetic data sets and real world images.