丛晓峰, 章军, 胡强. 基于对偶学习的图像去雾网络[J]. 应用光学, 2020, 41(1): 94-99. DOI: 10.5768/JAO202041.0102005
引用本文: 丛晓峰, 章军, 胡强. 基于对偶学习的图像去雾网络[J]. 应用光学, 2020, 41(1): 94-99. DOI: 10.5768/JAO202041.0102005
CONG Xiaofeng, ZHANG Jun, HU Qiang. Image defogging network based on dual learning[J]. Journal of Applied Optics, 2020, 41(1): 94-99. DOI: 10.5768/JAO202041.0102005
Citation: CONG Xiaofeng, ZHANG Jun, HU Qiang. Image defogging network based on dual learning[J]. Journal of Applied Optics, 2020, 41(1): 94-99. DOI: 10.5768/JAO202041.0102005

基于对偶学习的图像去雾网络

Image defogging network based on dual learning

  • 摘要: 针对光学成像设备在雾天搜集到的图像存在的降质问题,提出了一种基于对偶学习的从源域到目标域转换的对偶去雾网络以实现图像去雾功能。网络首先采用对偶生成对抗网络直接学习有雾图像与无雾图像之间的双向映射关系,并将有雾到无雾图像的映射作为初步的去雾结果,随后采用预训练模型在特征空间计算去雾图像与真实无雾图像的特征向量,运用欧式距离作为损失函数最小化特征向量之间的距离,以保证去雾图像在特征层面与真实无雾图像接近。实验结果表明,对偶去雾网络得到的去雾结果具有较高的峰值信噪比和较低的色差值,并能够有效保留图像的结构信息。

     

    Abstract: Aiming at the degradation problem of the images collected by the optical imaging equipment in hazy days, a Dual Dehazing Network based on dual learning from the source domain to the target domain is proposed to realize the image dehazing. First, the network learns the bilateral mapping relationship between the hazy image and the haze-free image by using the Dual Generative Adversarial Network, and obtains the preliminary dehazing result. Then the pre-training model is used to calculate the feature vector of the dehazed image and the real haze-free image in the feature space. The Euclidean distance is used as the loss function to minimize the distance between the feature vectors to ensure that the dehazed image is close to the real haze-free image at the feature level. The experimental results show that the dehazing results obtained by the Dual Dehazing network have higher peak signal-to-noise ratio and lower color difference, and can effectively preserve the structural information of the image.

     

/

返回文章
返回