All-optical diffractive neural networks for three-dimensional gesture recognition

Abstract

All-optical diffractive neural networks (D²NN) are a novel class of algorithmic models that leverage optical properties to overcome the limitations of traditional electronic neural networks, achieving higher computational efficiency and lower energy consumption. However, existing D²NN studies have primarily focused on objects located on a fixed plane. In real-world applications, targets are not stationary on a single plane, making current theoretical models difficult to apply in practice. To address this issue, this paper proposes a D²NN model operating at a wavelength of 1550 nm for targets distributed in three-dimensional space. Building upon the conventional model, the proposed approach incorporates the three-dimensional variations of targets encountered in practical scenarios, thereby enhancing the model’s robustness and applicability for spatially distributed objects. Experimental results show that when the object distance varies within the imaging range of 0.1 m to 1 m, the recognition accuracy of the unmodified model drops significantly once it exceeds the training plane. In contrast, although the improved model performs slightly worse than the traditional model on the specific plane, it maintains an overall recognition accuracy above 70% across the entire distance range. This demonstrates that the improved model proposed in this paper is theoretically more stable than the traditional one, providing a solid foundation for future practical applications.