Dual‐Wavelength Differential Diffractive Network for All‐Optical Object Classification

Abstract Optical computing, leveraging light propagation as engine, has emerged as a promising paradigm due to inherent advantages in parallel processing, high throughput, and energy‐efficient. However, Existing diffractive network architectures are typically designed for single‐wavelength illumination, limiting their ability to capture multiscale spatial‐frequency features of input objects. Here, a dual‐wavelength differential diffractive neural network () is proposed, which synergistically integrates complementary optical responses at two distinct wavelengths for all‐optical object classification. Each category of object is assigned to intensity signals corresponding to two distinct wavelengths, and classification inference is performed by maximizing the differential signal between these paired wavelengths. By leveraging end‐to‐end deep learning methods to jointly train differential diffractive networks, the resulting demonstrates a significant improvement in classification accuracy. Numerical validation shows that the four‐layer achieves 98.7% accuracy on MNIST and 90.1% on Fashion‐MNIST, surpassing traditional single‐wavelength methods by 11.4% and 7.8%, respectively. Furthermore, The framework is further extended to challenging scenarios involving diffusers along the optical path and demonstrated its superior robustness compared to traditional diffractive networks. The presented framework can be used to wide range of applications such as biomedical analysis and industrial inspection et al.