Scholarlink

<p>Computer-generated holography (CGH) faces the critical challenge of balancing computational efficiency with reconstruction quality in resource-constrained edge-computing environments. This study proposes Wave-H-Unet, a lightweight neural architecture that achieves real-time, high-quality hologram generation while maintaining deployability on edge devices. The framework integrates multi-domain analysis theory with lightweight network architecture design to establish an edge-computing-oriented hologram generation system. A core novel block featuring dual spatial-frequency-domain branches is developed, achieving cross-domain synergy that improves computational efficiency while maintaining balanced reconstruction quality. The frequency-domain branch employs a wavelet-based hybrid attention mechanism for global phase analysis, while the spatial-domain branch adopts depth-separable convolution for localized detail extraction, enabling complementary feature learning. For the frequency-domain branch, a unique module combining discrete wavelet transform is specifically designed and combined with attention mechanisms to form the wavelet hybrid attention module, effectively filtering redundant information while enhancing feature representation capabilities. Experimental results demonstrate that the proposed method achieves superior reconstruction quality with 25.15 dB average peak signal-to-noise ratio on the DIV2K high-resolution dataset, requiring only 0.39 million parameters to generate holograms within 0.4&#xa0;s. Compared with existing deep learning methods, the framework reduces floating-point operations by 74% and parameters by 46%. This work provides an end-to-end solution balancing efficiency and robustness for holographic systems on edge devices.</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

Wave-H-Unet: a lightweight neural network for real-time and high-quality computer-generated holography on edge devices

  • 期刊论文

  • 发表日期：

    2025年12月7日

查看全文
  • Xiaofei Nie,
  • Yudi Zhao,
  • Qiang He,
  • Kai Zhao

摘要

Computer-generated holography (CGH) faces the critical challenge of balancing computational efficiency with reconstruction quality in resource-constrained edge-computing environments. This study proposes Wave-H-Unet, a lightweight neural architecture that achieves real-time, high-quality hologram generation while maintaining deployability on edge devices. The framework integrates multi-domain analysis theory with lightweight network architecture design to establish an edge-computing-oriented hologram generation system. A core novel block featuring dual spatial-frequency-domain branches is developed, achieving cross-domain synergy that improves computational efficiency while maintaining balanced reconstruction quality. The frequency-domain branch employs a wavelet-based hybrid attention mechanism for global phase analysis, while the spatial-domain branch adopts depth-separable convolution for localized detail extraction, enabling complementary feature learning. For the frequency-domain branch, a unique module combining discrete wavelet transform is specifically designed and combined with attention mechanisms to form the wavelet hybrid attention module, effectively filtering redundant information while enhancing feature representation capabilities. Experimental results demonstrate that the proposed method achieves superior reconstruction quality with 25.15 dB average peak signal-to-noise ratio on the DIV2K high-resolution dataset, requiring only 0.39 million parameters to generate holograms within 0.4 s. Compared with existing deep learning methods, the framework reduces floating-point operations by 74% and parameters by 46%. This work provides an end-to-end solution balancing efficiency and robustness for holographic systems on edge devices.