Scholarlink

<p>Single image super-resolution (SISR) techniques leveraging deep neural networks have achieved remarkable improvements in output image quality. However, accurately reconstructing high-frequency components, especially edges, remains challenging. To address this issue, a Multi-Branch Generative Adversarial Network (MBGAN) has been designed for SISR with enhanced edge preservation. The MBGAN extracts thick and thin edge maps from the input image using first- and second-order derivative filters, respectively. These edge maps are then preserved and upscaled through two dedicated side branches within the network. The upscaled edge maps are fused with the structural map generated by the main branch, resulting in an output image with sharp edges and fine details. Furthermore, we employ distortion-oriented and perception-oriented loss functions to ensure accurate high-frequency information in the output image. Experimental results demonstrate the superiority of our proposed MBGAN over existing approaches, achieving state-of-the-art performance in terms of PSNR, SSIM, and LPIPS metrics.</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

A Multi-branch Generative Adversarial Network for Single Image Super-Resolution with Enhanced Edge Preservation

  • 期刊论文

  • 发表日期：

    2026年3月22日

查看全文
  • Khushboo Singla,
  • Rajoo Pandey,
  • Umesh Ghanekar

摘要

Single image super-resolution (SISR) techniques leveraging deep neural networks have achieved remarkable improvements in output image quality. However, accurately reconstructing high-frequency components, especially edges, remains challenging. To address this issue, a Multi-Branch Generative Adversarial Network (MBGAN) has been designed for SISR with enhanced edge preservation. The MBGAN extracts thick and thin edge maps from the input image using first- and second-order derivative filters, respectively. These edge maps are then preserved and upscaled through two dedicated side branches within the network. The upscaled edge maps are fused with the structural map generated by the main branch, resulting in an output image with sharp edges and fine details. Furthermore, we employ distortion-oriented and perception-oriented loss functions to ensure accurate high-frequency information in the output image. Experimental results demonstrate the superiority of our proposed MBGAN over existing approaches, achieving state-of-the-art performance in terms of PSNR, SSIM, and LPIPS metrics.