计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (24): 165-175.DOI: 10.3778/j.issn.1002-8331.2206-0117

• 图形图像处理 • 上一篇    下一篇

DRUSR:面向效果的图像超级分辨率重建

李昊,赵光哲   

  1. 北京建筑大学 电气与信息工程学院,北京 102616
  • 出版日期:2023-12-15 发布日期:2023-12-15

DRUSR:Effect-Oriented Super-Resolution Reconstruction of Images

LI Hao, ZHAO Guangzhe   

  1. School of Electrical and Information Engineering, Beijing University of Civil Engineering and Architecture, Beijing 102616, China
  • Online:2023-12-15 Published:2023-12-15

摘要: 图像超分辨率(super resolution,SR)重建是计算机视觉领域的热点问题,重点是利用深度学习将图像重建模型作为提高图像和视频分辨率的重要图像技术。往往这些重建模型的网络结构比较简单,就会导致梯度传递愈发困难,更会导致效率低下,而且重建后的图像依然存在细节丢失、噪声过大等问题,故提出一种改进生成网络中的残差块和判别网络中的判别模型的GAN(generative adversarial networks)图像超级分辨率重建模型。在模型结构上,将组成生成网络的基本单位简化成Conv+RELU,再进行重新设计,将密集残差网络的思想融入其中并重新设计组合成新的生成模型,将判别网络中的顺序连接的残差块进行了重新设计来实现更优的性能。训练模型所使用的数据集是DIV2K和Flickr2K。从最后得出的实验结果对比来看,在Set5、Set14、BSD100、Urban100四个公开的数据集上,所提出的模型相较于其他五个主流重建模型在图像重建质量的峰值信噪比(PSNR)、结构相似性(SSIM)上均提升1%~4%不等,在主观观感上也有所提高。

关键词: U型网络, 递减型残差块, 生成对抗网络, 超分辨率

Abstract: Super resolution(SR) reconstruction is a hot topic in the field of computer vision. This paper focuses on the detailed learning of image reconstruction models as an important image technology to improve image and video resolution. The network structure of these reconstructed models is relatively simple, which leads to lower transmission and lower efficiency. The reconstructed image has problems such as details and loss of excessive noise. This paper proposes a superresolution reconstruction model of GAN(generated adversarial networks) image which improves the generated residual blocks and generation discriminant network. The basic unit of the generation network is simplified and redesigned to Conv + RELU of the model structure. The idea of dense residual network is integrated into the model, the new generation model is redesigned, and the continuously connected residual block in the identification network is redesigned to achieve better performance. The datasets used to train models in this article are DIV2K and Flick2K. Compared with the other five main reconstruction models, the proposed model improves the peak signal to noise ratio (PSNR) and structural similarity (SSIM) of image reconstruction quality by 1% to 4% in four open datasets of Set5, Set14, BSD100, and Urban100. It also improves in subjective perception.

Key words: U network, decreasing residual block, generative adversarial networks(GAN), super-resolution