Scholarlink

<p>Single image super-resolution (SISR) is a fundamental task in image processing that aims to recover high-resolution images from low-resolution inputs. With the rapid development of deep learning, remarkable progress has been made in SISR. However, the substantial computational and memory costs of many advanced models limit their applicability on edge devices, making efficient super-resolution methods highly desirable. In this paper, we address efficient SISR by proposing a Multi-scale Partial Convolution Residual Distillation Network (MPCRDN), which is built upon the introduced Multi-scale Partial Convolution Residual Distillation Block (MPCRDB). Specifically, MPCRDB is carefully designed based on two key mechanisms: information distillation and multi-scale enhanced spatial attention. The former reduces channel redundancy and simplifies feature extraction, while the latter captures more accurate spatial information distributions with fewer parameters. Extensive experiments demonstrate that our model provides a favorable trade-off between reconstruction performance and model complexity compared to recently published baselines. Furthermore, to verify its practical efficiency for edge-device deployment, we evaluate the model on a mobile GPU (NVIDIA GeForce GTX 1650 Ti). The results indicate that MPCRDN achieves a highly competitive inference latency while maintaining excellent reconstruction quality.</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

Multi-scale partial convolution residual distillation network for efficient image super-resolution

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Chao Xie,
  • Fugui Xing,
  • Yi Zhu,
  • Shutong Ye

摘要

Single image super-resolution (SISR) is a fundamental task in image processing that aims to recover high-resolution images from low-resolution inputs. With the rapid development of deep learning, remarkable progress has been made in SISR. However, the substantial computational and memory costs of many advanced models limit their applicability on edge devices, making efficient super-resolution methods highly desirable. In this paper, we address efficient SISR by proposing a Multi-scale Partial Convolution Residual Distillation Network (MPCRDN), which is built upon the introduced Multi-scale Partial Convolution Residual Distillation Block (MPCRDB). Specifically, MPCRDB is carefully designed based on two key mechanisms: information distillation and multi-scale enhanced spatial attention. The former reduces channel redundancy and simplifies feature extraction, while the latter captures more accurate spatial information distributions with fewer parameters. Extensive experiments demonstrate that our model provides a favorable trade-off between reconstruction performance and model complexity compared to recently published baselines. Furthermore, to verify its practical efficiency for edge-device deployment, we evaluate the model on a mobile GPU (NVIDIA GeForce GTX 1650 Ti). The results indicate that MPCRDN achieves a highly competitive inference latency while maintaining excellent reconstruction quality.