Controlled numerical generation and reconstruction of off-axis digital holograms with object-specific frequency-domain filtering.

Deep-learning-based holographic reconstruction requires numerical data with reliable ground truth for training and quantitative evaluation. Experimental hologram acquisition and pixel-level phase annotation can be difficult to control, making physically defined numerical cases useful as reproducible benchmarks. Here, three analytically specified phase targets-a smooth Gaussian distribution, a four-level step distribution, and a radially structured Siemens Star-are used to generate off-axis digital holograms with known phase ground truth. A frequency-domain filtering procedure is applied before inverse Fourier reconstruction, followed by DCT-based global phase unwrapping. Reconstruction quality is quantified using root mean square error (RMSE), mean absolute error (MAE), correlation coefficient, and profile-based error analysis. The results show that reconstruction errors are comparatively small in smooth regions and increase near abrupt or high-spatial-frequency phase transitions. The three cases are therefore presented as controlled benchmark examples for algorithm verification and baseline comparison, not as a complete substitute for experimental holographic datasets or as a complete training corpus for deep-learning models.