Scholarlink

<p>The advent of deep learning has led to a notable improvement in super resolution (SR) performance in recent years. Because supervised deep learning SR methods often rely on training data generated by an assumed or predetermined degradation model, they perform exceptionally well in the ideal scenario, where the degradation model of a test low-resolution image does, in fact, conform with the assumed model (e.g., bicubic down-scaling) without unknown parameters like sensor noise, non-ideal point spread function (PSF), etc. However, this ideal setting is rarely suitable for actual low-resolution photographs. In this paper, we propose a hybrid method for the SR problem that considers both internal information to a given image and exterior information obtained by a pretrained SR network. Consequently, a hybrid network for SR is produced, which utilizes both information channels and adapts to the given image (and possibly its specific degradation). According to the experimental results on the standard dataset, the proposed approach has improved PSNR and SSIM. The maximum values of the two PSNR and SSIM measurements in the unknown downscaling kernel case are 29.50 and 0.93 respectively. The results of the experiments indicate that the proposed approach outperforms the state-of-the-art methods, especially when the degradation model is ambiguous or not ideal. Our source code is available at <a href="https://github.com/cvloc/FusionNet_SR">https://github.com/cvloc/FusionNet_SR</a>.</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

Internal–external networks for image quality improvement

  • 期刊论文

  • 发表日期：

    2025年9月26日

查看全文
  • Cu Vinh Loc,
  • Nguyen Thanh Hai,
  • Truong Xuan Viet,
  • Tran Hoang Viet,
  • Le Hoang Thao,
  • Nguyen Hoang Viet

摘要

The advent of deep learning has led to a notable improvement in super resolution (SR) performance in recent years. Because supervised deep learning SR methods often rely on training data generated by an assumed or predetermined degradation model, they perform exceptionally well in the ideal scenario, where the degradation model of a test low-resolution image does, in fact, conform with the assumed model (e.g., bicubic down-scaling) without unknown parameters like sensor noise, non-ideal point spread function (PSF), etc. However, this ideal setting is rarely suitable for actual low-resolution photographs. In this paper, we propose a hybrid method for the SR problem that considers both internal information to a given image and exterior information obtained by a pretrained SR network. Consequently, a hybrid network for SR is produced, which utilizes both information channels and adapts to the given image (and possibly its specific degradation). According to the experimental results on the standard dataset, the proposed approach has improved PSNR and SSIM. The maximum values of the two PSNR and SSIM measurements in the unknown downscaling kernel case are 29.50 and 0.93 respectively. The results of the experiments indicate that the proposed approach outperforms the state-of-the-art methods, especially when the degradation model is ambiguous or not ideal. Our source code is available at https://github.com/cvloc/FusionNet_SR.