计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (18): 230-238.DOI: 10.3778/j.issn.1002-8331.2306-0104

• 图形图像处理 • 上一篇    下一篇

具有自适应激活层和感受野块的超分辨网络

关宇珅,梁玉琦   

  1. 1.兰州交通大学 光电技术与智能控制教育部重点实验室,兰州 730070
    2.兰州交通大学 国家绿色镀膜技术与装备工程技术研究中心,兰州 730070
  • 出版日期:2024-09-15 发布日期:2024-09-13

Super-Resolution Network with Adaptive Activation Layer and Receptive Field Block

GUAN Yushen, LIANG Yuqi   

  1. 1.Key Laboratory of Optoelectronic Technology and Intelligent Control, Ministry of Education, Lanzhou Jiaotong University, Lanzhou 730070, China
    2.National Green Coating Technology and Equipment Engineering Technology Research Center, Lanzhou Jiaotong University, Lanzhou 730070, China
  • Online:2024-09-15 Published:2024-09-13

摘要: 基于生成对抗网络的超分辨网络在提升图像感知质量等方面获得了重大突破,解决了SR图像中的边缘平滑问题,但依然存在纹理细节缺失、噪点和伪影等问题。在此提出一种含有学习层的自适应激活层MetaAconc与残差网络重构结合,得到一种新的残差块ResMA。此外感受野块已经在目标检测等方面取得了不错的效果,为了增强特征细节,提出了增强感受野密集残差块(enhanced residual of receptive field dense block,ERRFDB)。将ResMA和ERRFDB进行网络结构重组,提出了一种全新的超分辨率生成器模型(SRRMA-RFB)。在SRGAN网络基础上,将生成器替换为SRRMA-RFB的网络模型称为SRGAN-ARF,为验证其重建图像的视觉效果和评价指标都有所提高,将其与SRGAN和ESRGAN算法进行对比。实验证明,所提算法在提高网络性能并且控制计算量的同时,使重建图像拥有更好的感知质量和纹理细节并且在减轻噪声方面具有一定的优势。

关键词: 图像超分辨重建, 自适应激活层, 感受野块

Abstract: The super-resolution network based on generative adversarial network has obtained a significant breakthrough in improving image perception quality. It solves the edge smoothing problem in SR images, but there are still problems such as missing texture details, noise and artifacts. Here, a new residual block ResMA is proposed by combining MetaAconc, an adaptive activation layer containing a learning layer, with the reconstruction of the residual network. In addition, the receptive field block has achieved good results in object detection, and in order to enhance the feature details, the enhanced residual of receptive field dense block (ERRFDB) is proposed. A new super-resolution generator model (SRRMA-RFB) is proposed by reorganizing the network structure of ResMA and ERRFDB. Based on the SRGAN network, the network model that replaces the generator with SRRMA-RFB is called SRGAN-ARF, and in order to verify that the visual effect and evaluation index of its reconstructed image are improved, it is compared with SRGAN and ESRGAN algorithms. Experiments show that the proposed algorithm not only improves the network performance and controls the amount of calculation, but also makes the reconstructed image have better perception quality and texture details and has certain advantages in reducing noise.

Key words: super-resolution image reconstruction, adaptive activation layer, receptive field block