Parallel Spatial–Frequency Diffractive Neural Networks for Autonomous-Driving Visual Understanding

Abstract

Diffractive neural networks (DNNs) offer a promising route toward high-throughput, low-power optical computing; however, their application to autonomous-driving perception remains limited by insufficient representational capacity for complex scenes, owing to their relatively simple architectures and reliance on spatial-domain information alone. Here, we propose a spatial–frequency diffractive neural network (SFDNN) architecture that jointly exploits spatial- and frequency-domain information through two parallel sub-DNN branches. By leveraging the complementarity of these two domains, the proposed architecture improves feature separability and enhances perception performance, thereby outperforming conventional DNNs in the autonomous-driving segmentation task. Numerical results show that, under the same modulation-layer constraint, the SFDNN consistently surpasses single-domain models in segmentation tasks, improving reconstruction fidelity and spatial overlap over the best single-domain baseline by 10.9% and 23.2%, respectively. The framework further demonstrates strong generalization to traffic sign classification, achieving accuracies of up to 93.67%. Experimentally, we validate the proposed framework using a hybrid optical–digital system and demonstrate reliable target-region segmentation in a physical optical setup. Furthermore, a micrometer-scale SFDNN chip fabricated by laser nanoprinting confirms the feasibility of on-chip integration. This work provides a practical pathway to efficient, intelligent optical perception for autonomous driving and related edge-vision applications.