Scholarlink

<p>Adversarial and periodic training method plays an essential role in video super-resolution, which can generate spatial high frequency detail and temporal consistency relation. However, this approach is based on unidirectional loop whose first frame can only use its own feature information, while the last frame can use all feature information of whole sequence. The biggest problem caused by this information imbalance is that early video frames are poorly reconstructed. To address these issues, we propose a novel video super-resolution model, called <b>Bi</b>directional <b>S</b>patio-<b>T</b>emporal <b>G</b>enerative <b>A</b>dversarial <b>N</b>etwork (Bi-STGAN), to generate fine detail and temporal consistency video by explicitly introducing the backward branch. Specifically, Bi-STGAN adopts an elaborately designed bidirectional branch structure so that the high-resolution frames estimated from front to back can be used as input for subsequent iterations from back to front. The advantage of Bi-STGAN is to enhance information gathering by utilizing information from past and future frames which can be cyclically passed through the time series. The experimental results show that compared with the baselines, Bi-STGAN achieves competitive improvement of 15.20% for LPIPS and 2.87dB for PSNR on the REDS4 dataset, thereby demonstrating the superiority of our state-of-the-art model on video super-resolution.</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

Bidirectional spatio-temporal generative adversarial network for video super-resolution

  • 期刊论文

  • 发表日期：

    2025年1月31日

查看全文
  • Peng Yang,
  • Zhangquan Chen,
  • Yuankang Sun,
  • Zhongjian Hu,
  • Bing Li

摘要

Adversarial and periodic training method plays an essential role in video super-resolution, which can generate spatial high frequency detail and temporal consistency relation. However, this approach is based on unidirectional loop whose first frame can only use its own feature information, while the last frame can use all feature information of whole sequence. The biggest problem caused by this information imbalance is that early video frames are poorly reconstructed. To address these issues, we propose a novel video super-resolution model, called Bidirectional Spatio-Temporal Generative Adversarial Network (Bi-STGAN), to generate fine detail and temporal consistency video by explicitly introducing the backward branch. Specifically, Bi-STGAN adopts an elaborately designed bidirectional branch structure so that the high-resolution frames estimated from front to back can be used as input for subsequent iterations from back to front. The advantage of Bi-STGAN is to enhance information gathering by utilizing information from past and future frames which can be cyclically passed through the time series. The experimental results show that compared with the baselines, Bi-STGAN achieves competitive improvement of 15.20% for LPIPS and 2.87dB for PSNR on the REDS4 dataset, thereby demonstrating the superiority of our state-of-the-art model on video super-resolution.