陈玮, 李正旺, 尹钟. 基于生成对抗网络的图像去雾算法[J]. 信息与控制, 2019, 48(6): 707-714, 722. DOI: 10.13976/j.cnki.xk.2019.9078
引用本文: 陈玮, 李正旺, 尹钟. 基于生成对抗网络的图像去雾算法[J]. 信息与控制, 2019, 48(6): 707-714, 722. DOI: 10.13976/j.cnki.xk.2019.9078
CHEN Wei, LI Zhengwang, YIN Zhong. Image Deblurring Algorithm Based on Generative Adversarial Network[J]. INFORMATION AND CONTROL, 2019, 48(6): 707-714, 722. DOI: 10.13976/j.cnki.xk.2019.9078
Citation: CHEN Wei, LI Zhengwang, YIN Zhong. Image Deblurring Algorithm Based on Generative Adversarial Network[J]. INFORMATION AND CONTROL, 2019, 48(6): 707-714, 722. DOI: 10.13976/j.cnki.xk.2019.9078

基于生成对抗网络的图像去雾算法

Image Deblurring Algorithm Based on Generative Adversarial Network

  • 摘要: 为了提高雾霾条件下拍摄到的模糊图像的质量,提出了一种基于生成对抗学习思想的卷积神经网络去雾算法.该卷积网络在生成模型部分将介质透射率和大气光值嵌入一个变量,进行了联合优化,避免了分离优化影响整体去雾性能的缺陷;在对抗模型部分,将生成模型部分的输出清晰图和真实的清晰图进行对比,从而判别生成的输出清晰图是否足够真实.为了生成更加逼真的清晰图像,引入了一种新的损失函数来优化网络参数,该损失函数同时纳入了图像的L2损失和双向梯度损失、特征损失和判别器损失,从4个方面来保证去雾性能的良好表现.除此之外,在训练网络的过程中,使用了真实图像和合成有雾图像同时作为数据集,其中的合成图像在合成过程中采用引导滤波算法,这样可以使得合成的有雾图像更加接近于自然情况.最后,引入了更多的评价指标验证了所提方法.基于不同方法的实验数据和实验结果证明了本文方法在已有方法上的提升.

     

    Abstract: To improve the quality of blurred images captured under haze conditions, we propose a convolutional network defogging algorithm based on the generative adversarial learning. The convolutional network jointly optimizes the dielectric transmission map and the atmospheric light value, which avoids the defect of separation optimization affecting the overall defogging performance. To generate more vivid and clear images, we introduce a new loss function to optimize the network. The loss function includes L2 loss and bidirectional gradient loss, feature loss, and discriminator loss. It guarantees good defogging performance from four aspects. In addition, in the network training process, we use real image and synthetic foggy images as datasets at the same time. In the synthesis process, we process the synthetic images using the guided filtering algorithm. Finally, we introduce several evaluation indexes to verify the effect of the method. The experimental data and results under different methods show that the method makes significant improvements on the basis of the existing methods.

     

/

返回文章
返回