计算机应用 ›› 2020, Vol. 40 ›› Issue (8): 2345-2350.DOI: 10.11772/j.issn.1001-9081.2019122142

• 虚拟现实与多媒体计算 • 上一篇    下一篇

基于球形矩匹配与特征判别的图像超分辨率重建

林静, 黄玉清, 李磊民   

  1. 西南科技大学 信息工程学院, 四川 绵阳 621010
  • 收稿日期:2019-12-23 修回日期:2020-03-29 出版日期:2020-08-10 发布日期:2020-04-23
  • 通讯作者: 黄玉清(1962-),女,四川绵阳人,教授,硕士,主要研究方向:图像处理、机器视觉、智能技术;hyq_851@163.com
  • 作者简介:林静(1993-),女,四川广安人,硕士研究生,主要研究方向:图像处理、机器视觉;李磊民(1960-),男,四川绵阳人,教授,硕士,主要研究方向:机器人控制、无线测控、图像处理、机器视觉。
  • 基金资助:
    国家自然科学基金面上项目(61673220)。

Image super-resolution reconstruction based on spherical moment matching and feature discrimination

LIN Jing, HUANG Yuqing, LI Leimin   

  1. School of Information Engineering, Southwest University of Science and Technology, Mianyang Sichuan 621010, China
  • Received:2019-12-23 Revised:2020-03-29 Online:2020-08-10 Published:2020-04-23
  • Supported by:
    This work is partially supported by the Surface Program of National Natural Science Foundation of China (61673220).

摘要: 由于网络训练不稳定,基于生成对抗网络(GAN)的图像超分辨率重建存在模式崩溃的现象。针对此问题,提出了一种基于球形几何矩匹配与特征判别的球面双判别器超分辨率重建网络SDSRGAN,通过引入几何矩匹配与高频特征判别来改善网络训练的稳定性。首先,生成器对图像提取特征并通过上采样生成重建图像;接着,球面判别器将图像特征映射至高维球面空间,充分利用特征数据的高阶统计信息;然后,在传统判别器的基础上增加特征判别器,提取图像高频特征,重建特征高频分量和结构分量两方面;最后,对生成器与双判别器进行博弈训练,提高生成器重建图像质量。实验结果表明,所提算法能有效收敛,其网络能够稳定训练,峰值信噪比(PSNR)为31.28 dB,结构相似性(SSIM)为0.872,而与双三次差值、超分辨率残差网络(SRResNet)、加速的卷积神经网络超分辨率(FSRCNN)、基于GAN的单图像超分辨率(SRGAN)和增强型超分辨率生成对抗网络(ESRGAN)算法相比,所提算法的重建图像具有更加逼真的结构纹理细节。所提算法为基于GAN的图像超分辨率研究提供了球形矩匹配与特征判别的双判别方法,在实际应用中可行且有效。

关键词: 生成对抗网络, 图像超分辨率重建, 高频特征, 双判别器, 模式崩溃

Abstract: Due to the instability of network training, the image super-resolution reconstruction based on Generative Adversarial Network (GAN) has a mode collapse phenomenon. To solve this problem, a Spherical double Discriminator Super-Resolution Generative Adversarial Network (SDSRGAN) based on spherical geometric moment matching and feature discrimination was proposed, and the stability of network training was improved by adopting geometric moment matching and discrimination of high-frequency features. First of all, the generator was used to produce a reconstructed image through feature extraction and upsampling. Second, the spherical discriminator was used to map image features to high-dimensional spherical space, so as to make full use of higher-order statistics of feature data. Third, a feature discriminator was added to the traditional discriminator to extract high-frequency features of the image, so as to reconstruct both the characteristic high-frequency component and the structural component. Finally, game training between the generator and double discriminator was carried out to improve the quality of the image reconstructed by the generator. Experimental results show that the proposed algorithm can effectively converge, its network can be stably trained, and has Peak Signal-to-Noise Ratio (PSNR) of 31.28 dB, Structural SIMilarity (SSIM) of 0.872. Compared with Bicubic, Super-Resolution Residual Network (SRResNet), Fast Super-Resolution Convolutional Neural Network (FSRCNN), Super-Resolution using a Generative Adversarial Network (SRGAN), and Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) algorithms, the reconstructed image of the proposed algorithm has more precise structural texture characteristics. The proposed algorithm provides a double discriminant method for spherical moment matching and feature discrimination for the research of image super-resolution based on GAN, which is feasible and effective in practical applications.

Key words: Generative Adversarial Network (GAN), image super-resolution reconstruction, high-frequency feature, double discriminator, mode collapse

中图分类号: