Complex-valued attention feature distillation network for high-fidelity phase-only hologram generation

Area of Science:

• Computer Vision
• Optics
• Artificial Intelligence

Background:

• Phase-only hologram (POH) generation is crucial for advanced holographic displays.
• Existing methods often struggle with computational efficiency and image fidelity.
• There is a need for advanced deep learning models to address these challenges.

Purpose of the Study:

• To propose a novel complex-valued attention feature distillation network (CAFDN) for efficient and high-fidelity POH generation.
• To introduce a lightweight module (CAFDN_Lite) for enhanced feature extraction and distillation.
• To optimize the balance between representational capacity and computational cost in holographic reconstruction.

Main Methods:

• Utilized dual U-Net architectures integrated with a novel CAFDN_Lite module.
• Implemented simplified upsampling and downsampling layers for computational efficiency.
• Employed hierarchical feature extraction with local attention, multi-scale analysis, and channel pruning within CAFDN_Lite.

Main Results:

• Achieved an average Peak Signal-to-Noise Ratio (PSNR) of 32.52 dB and Structural Similarity Index (SSIM) of 0.861.
• Demonstrated a fast running time of 36 ms, outperforming conventional methods.
• Verified superior detail reproduction and image quality through numerical and optical experiments.

Conclusions:

• The proposed CAFDN framework offers significant advancements in POH generation.
• The network achieves a superior balance between image quality and computational efficiency.
• The method shows strong potential for real-time, high-fidelity holographic display applications.