Single Image Super-Resolution Reconstruction based on the ResNeXt Network

  • PDF / 919,454 Bytes
  • 12 Pages / 439.37 x 666.142 pts Page_size
  • 75 Downloads / 233 Views

DOWNLOAD

REPORT


Single Image Super-Resolution Reconstruction based on the ResNeXt Network Fangzhe Nan 1 & Qingliang Zeng 1 & Yanni Xing 1 & Yurong Qian 1 Received: 12 August 2019 / Revised: 31 January 2020 / Accepted: 7 May 2020 # Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract

To solve the complex computation, unstable network and slow learning speed problems of a generative adversarial network for image super-resolution (SRGAN), we proposed a single image super-resolution reconstruction model called the Res_WGAN based on ResNeXt. The generator is constructed by the ResNeXt network, which reduced the computational complexity of the model generator to 1/8 that of the SRGAN. The discriminator was constructed by the Wasserstein GAN(WGAN), which solved the SRGAN’s instability. By removing the normalization operation in the residual network, the learning rate is improved. The experimental results from the Res_WGAN demonstrated that the proposed model achieved better performance in the subjective and objective evaluations using four public data sets compared with other state-of-the-art models. Keywords Single image super-resolution reconstruction . ResNeXt . WGAN . Deep learning

1 Introduction Single image super-resolution reconstruction (SISR) refers to the reconstruction of corresponding high-resolution images from a single low-resolution image [23]. This technology plays an extremely important role in many image-related applications such as medical image processing

* Yurong Qian [email protected] Fangzhe Nan [email protected] Qingliang Zeng [email protected] Yanni Xing [email protected]

1

College of Software, Xinjiang University, Urumqi 830046, China

Multimedia Tools and Applications

[19], remote sensing [15], and video surveillance [16]. In recent years, due to the breakthrough progress in deep learning in other computer vision fields, people have tried to introduce deep learning models to solve image super-resolution reconstruction problems via end-to-end training. Convolutional neural networks (CNNs [12]), recurrent neural networks (RNNs [4]), residual networks (ResNets [9]), dense residual networks (DenseNets [5]), and generative adversarial networks (GANs [6]), such as the SRCNN [2], DRRN [24], IRCNN [8], SRDenseNet [20], and SRGAN [13], have been used to solve the SISR problem. However, most of the current deep-learning-based SISR models have the following two problems. (1) In a convolutional neural network, the model performance is usually enhanced by stacking multiple layers or increasing the number of filters. For example, the VDSR [11] used residual learning and adjustable gradient clipping to accelerate convergence. The DRCN [22] used residual learning to increase the network depth and improve the network performance. The DRRN was inspired by the ResNet, VDSR and DRCN and adopted a deeper network structure to improve the performance. However, this kind of method increases the computational complexity and memory consumption by expanding the convolution kernel size or increasing the layer de