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 (4): 165-172. doi: 10.19678/j.issn.1000-3428.0060818

• 图形图像处理 • 上一篇    下一篇

基于条件生成对抗网络与知识蒸馏的单幅图像去雾方法

何涛1, 俞舒曼1, 徐鹤2,3   

  1. 1. 南京邮电大学 电子与光学工程学院、微电子学院, 南京 210023;
    2. 南京邮电大学 计算机学院、软件学院、网络空间安全学院, 南京 210023;
    3. 江苏省无线传感网高技术研究重点实验室, 南京 210023
  • 收稿日期:2021-02-05 修回日期:2021-05-06 发布日期:2021-05-08
  • 作者简介:何涛(1972—),男,讲师、博士,主研方向为无线传感器网络、移动认知无线电网络;俞舒曼,硕士研究生;徐鹤,副教授、博士。
  • 基金资助:
    国家重点研发计划(2019YFB2103003);江苏省科技支撑计划项目(BE2019740);江苏省高等学校自然科学研究重大项目(18KJA520008);江苏省“六大人才高峰”高层次人才项目(RJFW-111)。

Single Image Dehazing Method Based on Conditional Generative Adversarial Network and Knowledge Distillation

HE Tao1, YU Shuman1, XU He2,3   

  1. 1. College of Electronic and Optical Engineering, College of Microelectronics, Nanjing University of Posts and Telecommunications, Nanjing 210023, China;
    2. College of Computer, College of Software, College of Cyberspace Security, Nanjing University of Posts and Telecommunications, Nanjing 210023, China;
    3. Jiangsu Key Laboratory of Wireless Sensor Network High Technology Research, Nanjing 210023, China
  • Received:2021-02-05 Revised:2021-05-06 Published:2021-05-08

摘要: 生成对抗网络广泛应用于图像去雾领域,但通常需要较大的计算量和存储空间,从而限制了其在移动设备上的应用。针对该问题,提出一种基于条件生成对抗网络与知识蒸馏的去雾方法KD-GAN。将频率信息作为去雾的附加约束条件,通过傅里叶变换、拉普拉斯算子、高斯滤波器分别滤除原始图像的高频或低频信息,生成对应的高频和低频图像,并将融合得到的图像作为判别器的输入,以改进雾天图像的去雾效果。在此基础上,将原重型教师网络的知识迁移到具有较少权值参数的轻量型学生网络生成器中,并对轻量型学生网络进行训练,使其以更快的收敛速度达到与教师网络相近的去雾性能。在OTS和HSTS数据集上的实验结果验证了该方法的有效性,在学生网络的参数规模仅为教师网络1/2的条件下,学生网络在迭代第3×104次时,生成器输出图像的峰值信噪比和结构相似性已接近于教师网络迭代第5×104次时的数值,训练速度加快了约1.67倍。

关键词: 图像去雾, 生成对抗网络, 知识蒸馏, 教师网络, 学生网络

Abstract: The Generative Adversarial Network(GAN) is widely used in the field of image dehazing;however it usually requires extensive computations and a large amount storage space, limiting its application in mobile devices.To solve this problem, this study proposes a dehazing method based on knowledge distillation and conditional generative adversarial network (KD-GAN).Taking the frequency information as the additional constraint condition for dehazing, the high-frequency and low-frequency information of the original image are filtered using the Fourier transform, Laplace operator, and Gaussian filter to generate the corresponding high-frequency and low-frequency images, whereby the fused image is used as the input of the discriminator to further improve the dehazing of the fog image.In the process, the knowledge of the original heavy teacher network is transferred to the lightweight student network generator with fewer weight parameters, and the lightweight student network is trained to achieve a dehazing performance similar to that of the teacher network with faster convergence speed.The experimental results on the OTS and HSTS datasets verify the effectiveness of this method.Under the condition that the parameter scale of the student network is only half that of the teacher network, the student network is iterated 3×104 times.The peak signal-to-noise ratio and structural similarity of the output image of the generator are close to those of the teacher network iterated 5×104 times, while the training speed is about 1.67 times faster than that of the teacher network.

Key words: image dehazing, Generative Adversarial Network(GAN), knowledge distillation, teacher network, student network

中图分类号: