Scholarlink

<p>In distributed video compressed sensing tasks, the intrinsic inter-frame and intra-frame correlations, along with sparsity within video sequences, are often underutilized. While deformable 3D convolutions effectively capture spatiotemporal information for video restoration, their large parameter size leads to high computational complexity, posing significant challenges for training and deployment. Focusing on these issues, we proposed a lightweight Deformable 3D Convolution applied to distributed video compressed sensing network (CSLD3D), which optimizes the model structure by constructing two branching networks for key and non-key frames in the initial reconstruction stage of each frame, while the deep reconstruction stage utilizes lightweight 3D deformable convolution to achieve further reconstruction. Extensive experiments on six standard CIF video sequences confirm that the proposed method consistently surpasses state-of-the-art approaches. Specifically, at a non-key-frame measurement rate of 0.1, CSLD3D attains an average PSNR of 37.45&#xa0;dB and SSIM of 0.9476, representing a 0.90&#xa0;dB improvement in PSNR over the recent WRDD method while reducing the parameter count by approximately 60%. These results demonstrate that CSLD3D achieves superior reconstruction quality with substantially lower model complexity. Code and results will be available at: <a href="https://github.com/FreshChance/LightweightDVCSNet.">https://github.com/FreshChance/LightweightDVCSNet.</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

CSLD3D: A lightweight distributed video compressed sensing network based on lightweight deformable 3D convolution

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Xing Gao,
  • Jinyao Zhu,
  • Xiuli Du,
  • Yana Lv,
  • Zhiguo Liu

摘要

In distributed video compressed sensing tasks, the intrinsic inter-frame and intra-frame correlations, along with sparsity within video sequences, are often underutilized. While deformable 3D convolutions effectively capture spatiotemporal information for video restoration, their large parameter size leads to high computational complexity, posing significant challenges for training and deployment. Focusing on these issues, we proposed a lightweight Deformable 3D Convolution applied to distributed video compressed sensing network (CSLD3D), which optimizes the model structure by constructing two branching networks for key and non-key frames in the initial reconstruction stage of each frame, while the deep reconstruction stage utilizes lightweight 3D deformable convolution to achieve further reconstruction. Extensive experiments on six standard CIF video sequences confirm that the proposed method consistently surpasses state-of-the-art approaches. Specifically, at a non-key-frame measurement rate of 0.1, CSLD3D attains an average PSNR of 37.45 dB and SSIM of 0.9476, representing a 0.90 dB improvement in PSNR over the recent WRDD method while reducing the parameter count by approximately 60%. These results demonstrate that CSLD3D achieves superior reconstruction quality with substantially lower model complexity. Code and results will be available at: https://github.com/FreshChance/LightweightDVCSNet.