作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (3): 280-287. doi: 10.19678/j.issn.1000-3428.0059926

• 图形图像处理 • 上一篇    下一篇

基于循环生成对抗网络的图像去雾算法

郭梦琰1, 张娟1, 刘巧红2, 蔡立志3   

  1. 1. 上海工程技术大学 电子电气工程学院, 上海 201620;
    2. 上海健康医学院 医疗器械学院, 上海 201318;
    3. 上海计算机软件技术开发中心, 上海 200235
  • 收稿日期:2020-11-06 修回日期:2021-01-10 发布日期:2022-03-11
  • 作者简介:郭梦琰(1993-),女,硕士研究生,主研方向为计算机视觉、图像处理;张娟、刘巧红,副教授、博士;蔡立志,研究员、博士。
  • 基金资助:
    国家自然科学基金(61801288)。

Image Dehazing Algorithm Based on Recurrent Generative Adversarial Network

GUO Mengyan1, ZHANG Juan1, LIU Qiaohong2, CAI Lizhi3   

  1. 1. School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China;
    2. School of Medical Devices, Shanghai University of Medicine and Health Sciences, Shanghai 201318, China;
    3. Shanghai Computer Software Technology Development Center, Shanghai 200235, China
  • Received:2020-11-06 Revised:2021-01-10 Published:2022-03-11

摘要: 大气散射模型与有雾图像及对应清晰图像间的映射模型不适配,导致使用大气散射模型进行图像去雾处理时,图像存在颜色偏差、纹理细节粗糙等问题。基于模拟生物视觉系统的反馈原理,提出一种端到端的循环生成对抗网络算法,以解决误差累积造成的去雾图像低质的问题。通过生成模块将循环神经网络的隐藏状态作为反馈信息,以指导低级模糊特征信息生成更加丰富的高级特征。循环结构能够保证先前的网络层可以使用到后面网络层的高级特征信息,从而减少误差累积。此外,该算法能够根据判别模块的损失来评估重建图像的质量。实验结果表明,与GCANet算法相比,所提算法在SOTS测试集上的平均峰值信噪比和结构相似性,在室内分别提升3.41%和0.57%,在室外分别提升3.48%和1.39%,且在真实世界的数据集上进行图像去雾后,在视觉上避免了颜色失真和光晕问题。

关键词: 图像去雾, 循环卷积神经网络, 生成对抗网络, 编码-解码模式, 反馈连接

Abstract: In image dehazing tasks, sometimes the atmospheric scattering model does not match the model that maps hazy images to clear images, as a result, there are some problems in image defogging using atmospheric scattering model, such as color deviation, rough texture details and so on.To address the accumulated errors and improve the quality of dehazed images, this paper proposes an image dehazing algorithm using an end-to-end recurrent generative adversarial network by simulating the feedback mechanism of biological visual systems.The hidden state of the recurrent neural network in the generative module acts as feedback information to guide the low-level hazy feature information to generate richer high-level features.The recurrent structure ensures that the previous network layer can access the high-level feature information of the next network layer, thus reducing error accumulation.Finally, the quality of reconstructed image is judged according to the loss of the discriminative module.Experimental results on the SOTS test set show that compared with the GCANet-based algorithm, the proposed algorithm provides a 3.41% increase in average Peak Signal-to-Noise Ratio (PSNR) and a 0.57% increase in Structural Similarity (SSIM) in the indoor environment.For the outdoor environment, the proposed algorithm provides a 3.48% increase in PSNR and a 1.39% increase in SSIM.After image defogging on real-world data sets, the problems of color distortion and halo are avoided visually.

Key words: image dehazing, Recurrent Neural Network(RNN), Generative Adversarial Network (GAN), encoding-decoding mode, feedback connection

中图分类号: