一种改进的SRGAN图像超分辨重建算法

2021-06-08 11:58:30乔昕魏延
计算机时代 2021年1期

乔昕 魏延

摘  要: 现存的图像超分辨率重建算法存在模型训练不稳定、参数量多、模型收敛速度慢等缺点。在生成对抗网络的超分辨率算法(SRGAN)基础上,将轻量化的密集连接网络(DenseNet)作为生成对抗网络的生成器,使用WGan对判别器进行优化,利用Wasserstein代替SRGAN的JS散度,使其能够在网络参数更少、计算量更小的基础上实现更优的性能。实验结果表明,在四个公开的数据集上,所提出的模型比较主流重建模型在图像重建质量的峰值信噪比(PSNR)和结构相似性(SSIM)两个客观指标和主观视觉效果上都有所提高。

关键词: 图像重建; 轻量化; 密集连接网络; 生成对抗网络

中图分类号:TP183;TP391.41          文献标识码:A     文章编号:1006-8228(2021)01-72-04

Research on the algorithm of image super-resolution reconstruction with improved SRGAN

Qiao Xin, Wei Yan

(Chongqing Normal University College of Computer and Information Science, Chongqing 401331, China)

Abstract: The existing image super-resolution reconstruction algorithms have the disadvantages of model training instability, large parameter stakes and slow model convergence speed. On the basis of the Super-Resolution Generative Adversarial Networks (SRGAN), the lightweight Densely Connected Convolutional Networks (DenseNet) is used as the generator to generate the GAN, WGan is used to optimize the distinguisher, and the Wasserstein is used instead of JS dispersion of SRGAN, so that it can realize better performance with fewer network parameters and less computation. The experimental results show that, on the four exposed data sets, the proposed model improved the two objective indicators and subjective visual effects of the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) of image reconstruction quality compared with the mainstream reconstruction models.

Key words: image reconstruction; lightweight; Densely Connected Convolutional Networks; Generative Adversarial Networks

0 引言

現如今,图像作为一种极其重要的信息载体,在医学、军民卫星图像和安保视频检测监控等应用领域中对其质量的要求越来越严格,但是图像会由于气象环境和采集设备的影响,易导致采集的图像出现质量低或者细节的纹理信息缺乏等问题。图像超分辨率重建[1](Super-Resolution,SR)就是将低分辨率图像重建为对应的高分辨率图像。如今,图像超分辨率技术也逐渐成为计算机视觉领域的研究热点。

最近,随着深度学习的迅猛发展,深度学习模型开始被应用于图像重建这一领域。如:Dong等人在2014年成功提出了的一种基于卷积神经网络的超分辨率重建算法模型SRCNN[1](Super-Resolution using Convolutional Neural Networks)。2016年,又继续对SRCNN进行了改进,提出FSRCNN[2](Fast Super-Resolution Convolutional Neural Networks),不需要在网络外进行放大图像尺寸,用小卷积代替大卷积,在训练速度和图像质量上均有所提高。……

登录APP查看全文