Scholarlink

Abstract <p>Image superresolution reconstruction methods based on convolutional neural network is confronting the problems of single scale feature extraction scale and small receptive field. In order to solve these problems, an improved method of single image superresolution reconstruction is proposed based on multiscale feature extraction and hole residual network. In the initial stage of the network, the multiscale feature extraction module is used to improve the problem of insufficient extraction of high-frequency useful information of the image; in the middle stage of the network, six residual blocks with different expansion rates are connected in series to increase the receptive field, eliminate the “gridding” effect, and improve the transmission efficiency of image detail information. The experimental results on Set5, Set14, and BSD100 data sets show that the reconstruction results of this method are better than those of the other three algorithms. Compared with VDSR method, the average PSNR is increased by 0.73 dB and the average SSIM is increased by 0.0213; compared with Bicubic method, the average PSNR increased by 2.42 dB and the average SSIM increased by 0.0899.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Image Superresolution Reconstruction Method Research Based on Improved Residual Network

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Li Lan,
  • Zhang Yun,
  • Lin Guo Liang,
  • Xu Cheng Jun

摘要

Abstract

Image superresolution reconstruction methods based on convolutional neural network is confronting the problems of single scale feature extraction scale and small receptive field. In order to solve these problems, an improved method of single image superresolution reconstruction is proposed based on multiscale feature extraction and hole residual network. In the initial stage of the network, the multiscale feature extraction module is used to improve the problem of insufficient extraction of high-frequency useful information of the image; in the middle stage of the network, six residual blocks with different expansion rates are connected in series to increase the receptive field, eliminate the “gridding” effect, and improve the transmission efficiency of image detail information. The experimental results on Set5, Set14, and BSD100 data sets show that the reconstruction results of this method are better than those of the other three algorithms. Compared with VDSR method, the average PSNR is increased by 0.73 dB and the average SSIM is increased by 0.0213; compared with Bicubic method, the average PSNR increased by 2.42 dB and the average SSIM increased by 0.0899.