Toward high-quality hologram generation via a complex-valued efficient hybrid attention network

Area of Science:

• Computer Vision
• Holography
• Deep Learning

Background:

• Traditional U-Net methods for computer-generated holography (CGH) struggle with computational efficiency and feature extraction.
• This limits both the quality of hologram reconstruction and processing speed, hindering real-time applications.

Purpose of the Study:

• To propose a novel complex-valued efficient hybrid attention network (CEHAN) for high-quality hologram generation.
• To enhance computational resource efficiency and feature extraction capabilities in deep learning-based CGH.

Main Methods:

• Developed a CEHAN architecture with two sub-networks: complex amplitude inference network (CAIN) and hologram encoding network (HEN).
• Incorporated a complex efficient attention (CEA) mechanism in the downsampling module for improved accuracy and efficiency.
• Integrated a hybrid attention block (HAB) combining channel and spatial attention for optimized feature extraction.

Main Results:

• Achieved a processing time of 16 ms per frame.
• Attained an average Peak Signal-to-Noise Ratio (PSNR) of 35.71 dB and a Structural Similarity Index Measure (SSIM) of 0.944 on the DIV2K dataset.
• Demonstrated superior detail reproduction and image quality compared to conventional methods in simulations and optical experiments.

Conclusions:

• The proposed CEHAN method offers significant improvements in both reconstruction quality and computational efficiency for CGH.
• The framework shows strong potential for practical applications in holographic displays due to reduced computational demands and enhanced performance.
• CEHAN addresses the limitations of U-Net-based approaches, paving the way for more efficient and effective holographic technologies.