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 ›› 2021, Vol. 47 ›› Issue (3): 249-255. doi: 10.19678/j.issn.1000-3428.0057168

• 图形图像处理 • 上一篇    下一篇

基于生成式对抗网络的图像超分辨率重建算法

姜玉宁1, 李劲华1,2, 赵俊莉1   

  1. 1. 青岛大学 数据科学与软件工程学院, 山东 青岛 266071;
    2. 青岛市船舶建造智能精度测量工程研究中心, 山东 青岛 266071
  • 收稿日期:2020-01-10 修回日期:2020-02-24 发布日期:2021-03-15
  • 作者简介:姜玉宁(1993-),女,硕士研究生,主研方向为计算机图形学、计算机视觉;李劲华(通信作者),教授、博士;赵俊莉,副教授、博士。
  • 基金资助:
    国家自然科学基金(61702293);山东省重点研发计划重大科技创新工程项目(2019JZZY020101)。

Image Super-Resolution Reconstruction Algorithm Based on Generative Adversarial Networks

JIANG Yuning1, LI Jinhua1,2, ZHAO Junli1   

  1. 1. School of Data Science and Software Engineering, Qingdao University, Qingdao, Shandong 266071, China;
    2. Qingdao Shipbuilding Intelligent Precision Measurement Engineering Research Center, Qingdao, Shandong 266071, China
  • Received:2020-01-10 Revised:2020-02-24 Published:2021-03-15

摘要: 针对传统图像超分辨率重建算法存在网络训练困难与生成图像存在伪影的问题,提出一种利用生成式对抗网络的超分辨率重建算法。去除生成式对抗网络的批量归一化层降低计算复杂度,将其中的残差块替换为密集残差块构成生成网络,使用VGG19网络作为判别网络的基础框架,以全局平均池化代替全连接层防止过拟合,引入纹理损失函数、感知损失函数、对抗损失函数和内容损失函数构成生成器的总目标函数,利用纹理损失增强局部信息匹配度,采用激活层前的特征信息计算感知损失获取更多细节特征,使用WGAN-GP理论优化网络模型的对抗损失加速收敛,运用内容损失提升图像低频信息的准确性。实验结果表明,该算法重建图像的平均峰值信噪比为27.97 dB,平均结构相似性为0.777,与SRGAN和EDSR等算法相比,其在未延长较多运行时间的情况下,重建结果的纹理细节更清晰且亮度信息更准确,更符合视觉感官评价要求。

关键词: 超分辨率重建, 生成式对抗网络, 密集卷积网络, 纹理损失, 梯度惩罚Wasserstein生成式对抗网络

Abstract: The existing image Super-Resolution (SR) reconstruction algorithms have difficulty in network training and cause artifacts in the generated images.To address the problem,this paper proposes a SR reconstruction algorithm based on Generative Adversarial Networks (GAN).The Batch Normalization(BN)layer of the Generative Adversarial Networks(SRGAN)is removed to reduce the computational complexity,and the Residual Block (RB) is replaced by Residual Dense Block (RDB) to form the generative network.VGG19 network is used as the basic framework of the discrimination network.The Global Average Pooling (GAP)is used to replace the full connection layer to prevent over fitting.Texture loss function, perceptual loss function, adversarial loss function and content loss function are introduced to form the overall objective function of the generator. Texture loss is used to enhance the matching degree of local information, and the feature information in front of the activation layer is used to calculate the perceptual loss to obtain more detailed features. Wasserstein Generative Adversarial Nets-Gradient Penalty(WGAN-GP) theory is used to optimize the adversarial loss of the network model to accelerate the convergence, and content loss is used to improve the accuracy of low-frequency information of the image.Experimental results show that the average Peak Signal to Noise Ratio (PSNR) of the image reconstructed by the proposed algorithm is 27.97 dB,and its average Structural Similarity(SSIM)is 0.777.Compared with SRGAN and EDSR algorithms,the proposed algorithm improves the texture details and brightness of the reconstructed image,making it more in line with the requirements of visual sensory evaluation without prolonging much running time.

Key words: Super-Resolution(SR) reconstruction, Generative Adversarial Networks(GAN), Dense Convolutional Networks(DenseNet), texture loss, Wasserstein Generative Adversarial Nets-Gradient Penalty (WGAN-GP)

中图分类号: