Noise-robust optical edge detection using diffractive deep neural networks

Optical edge detection is essential for high-speed, energy-efficient imaging systems in applications such as autonomous driving, machine vision, and biomedical imaging. Traditional optical edge detection methods, such as 4-f systems, are limited to fixed linear filtering and lack robustness in noisy or complex environments. To overcome these limitations, we propose a noise-robust optical edge detection framework based on Diffractive Deep Neural Networks (D2NNs). The system comprises cascaded trainable phase-only diffractive layers, optimized via end-to-end learning to perform edge detection and suppress noise directly during optical propagation. As a proof of concept, we implement a three-layer D2NNs trained on digit and letter datasets with synthetic noise. An edge-enhanced weighted Mean Squared Error (MSE) loss is used to guide the network toward accurate edge localization. The D2NNs effectively extracts edge contours and suppresses artifacts, even when the input suffers from severe noise (Structural Similarity Index Measure (SSIM) = 0.65). Unlike traditional fixed architectures, the D2NNs supports reconfigurable depth, enabling scalable and energy-efficient optical processing. These results highlight the potential of D2NNs as a compact, adaptive, and high-performance platform for all-optical edge detection and computational imaging.