High-Quality Phase-Only Holography via Frequency-Domain Hybrid Encoding and Multi-Scale Complex Attention

Holography is a key technology in 3D display and optical imaging. Phase-only spatial light modulators have emerged as the core implementation platform for computer-generated holography due to their excellent diffraction efficiency and reconstruction quality. However, most existing deep learning-based CGH methods extract features by enlarging the receptive field of complex-valued convolutions in the spatial domain, which increases computational cost while failing to explicitly model the frequency-domain characteristics of optical wavefields. To address this limitation, we propose an end-to-end neural network FMCCA-Net for high-quality phase-only CGH generation. The proposed network integrates complex-valued operations with a hybrid frequency-domain encoding strategy and a multi-scale convolutional attention mechanism. By explicitly extracting frequency-domain features and fusing them with spatial-domain representations, the proposed method effectively enhances fine details in reconstructed holographic images. Comprehensive numerical simulations and optical experiments are conducted on the DIV2K dataset. Experimental results demonstrate that FMCCA-Net can reconstruct holograms with a resolution of 1920×1072 in approximately 15 ms, achieving an average peak signal-to-noise ratio of 34.69 dB and an average structural similarity index of 0.92. These results verify that the proposed method can stably generate holograms with details and well-suppressed artifacts, providing an efficient solution for practical high-quality holographic display applications.