Addressing scanning and imaging distortions in decryption of secured computer-generated holograms using convolutional neural networks

We recently introduced a novel method for securing valuable documents by integrating a random encryption layer into computer-generated holograms (CGHs) using the XOR encryption technique. This approach enhances security by making it significantly more difficult to replicate or counterfeit the information of secured holographic patterns or CGH. However, traditional decryption methods rely on the digital capturing process of the CGH, a process that is highly sensitive to distortions introduced during scanning or imaging. These distortions arise from several factors, including environmental conditions (e.g., dust, dirt, or smudges), variations in scanning angle and distance, non-uniform illumination, and limitations of the imaging system, such as camera resolution and autofocus inaccuracies. Such distortions reduce the fidelity of the reconstructed holographic data, thereby compromising the decryption process. To address these challenges, we propose a compensation framework based on convolutional neural networks (CNNs). The CNN is trained as a black-box model to detect and correct distortions by learning the differences between the original, numerically designed CGH and its scanned counterparts that involve distortions. The corrected CGH pattern is then used for accurate decryption. To validate the effectiveness of the proposed approach, we assess the reconstructed CGH using key performance metrics, including peak signal-to-noise ratio (PSNR) and mean square error (MSE). Experimental results demonstrate a marked improvement, with the CNN-enhanced CGH achieving a significant increase in PSNR of 4.9 dB and comparable diffraction efficiency to the original design. These results highlight the potential of deep learning-based methods in overcoming imaging distortions and improving the robustness and reliability of CGH-based security systems.