Scholarlink

<p>Efficient image super-resolution (SR) remains important for deploying high-quality image reconstruction on resource-constrained multimedia devices. However, many existing lightweight SR models still face difficulties in balancing receptive field expansion and computational efficiency, which limits their capability for multi-scale spatial representation. To address this issue, we propose the Progressive Multi-branch Distillation Attention Network (PMDAN), a lightweight SR framework based on a progressive dilated feature extraction strategy. The proposed Progressive Dilated Feature Extraction (PDFE) module incorporates Multi-branch Dilated Convolutions (MDC) within a Residual Multi-branch Dilated Block (R-MDB), enabling the network to effectively enlarge the effective receptive field while maintaining low computational complexity. In addition, we introduce the Scalable Multi-branch Dilated Attention (SMDA) mechanism, which integrates multi-scale dilated spatial modeling with local feature interaction to improve cross-scale feature aggregation. Extensive experiments on several benchmark datasets demonstrate that PMDAN achieves competitive reconstruction performance among recent lightweight SR methods. Compared with the representative lightweight model RFDN, PMDAN achieves improved reconstruction quality while reducing the number of parameters by approximately 24% and Multi-Adds by 5%. Further analyses show that the proposed progressive multi-branch design contributes to improved spatial dependency modeling and detail reconstruction under lightweight constraints, indicating its potential applicability in practical multimedia image restoration scenarios. The source code is available at <a href="https://github.com/nameycw2/PMDAN">https://github.com/nameycw2/PMDAN</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

Pmdan: progressive multi-branch distillation attention network for efficient image super-resolution

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Siwei Wei,
  • Changwen Yin,
  • Xueli Chang,
  • Ruoxi Wang,
  • Chunzhi Wang

摘要

Efficient image super-resolution (SR) remains important for deploying high-quality image reconstruction on resource-constrained multimedia devices. However, many existing lightweight SR models still face difficulties in balancing receptive field expansion and computational efficiency, which limits their capability for multi-scale spatial representation. To address this issue, we propose the Progressive Multi-branch Distillation Attention Network (PMDAN), a lightweight SR framework based on a progressive dilated feature extraction strategy. The proposed Progressive Dilated Feature Extraction (PDFE) module incorporates Multi-branch Dilated Convolutions (MDC) within a Residual Multi-branch Dilated Block (R-MDB), enabling the network to effectively enlarge the effective receptive field while maintaining low computational complexity. In addition, we introduce the Scalable Multi-branch Dilated Attention (SMDA) mechanism, which integrates multi-scale dilated spatial modeling with local feature interaction to improve cross-scale feature aggregation. Extensive experiments on several benchmark datasets demonstrate that PMDAN achieves competitive reconstruction performance among recent lightweight SR methods. Compared with the representative lightweight model RFDN, PMDAN achieves improved reconstruction quality while reducing the number of parameters by approximately 24% and Multi-Adds by 5%. Further analyses show that the proposed progressive multi-branch design contributes to improved spatial dependency modeling and detail reconstruction under lightweight constraints, indicating its potential applicability in practical multimedia image restoration scenarios. The source code is available at https://github.com/nameycw2/PMDAN.