image: Figure 1: Schematic illustration of Anti-Interference Diffractive Deep Neural Network. The network can classify handwritten digits (0-5) in multi-object scenarios, including intra-class interference, inter-class interference, and dynamic interference.
Credit: Nan Zhang et al.
Optical neural networks (ONNs) are emerging as a promising neuromorphic computing paradigm for object recognition, offering unprecedented advantages in light-speed computation, ultra-low power consumption, and inherent parallelism. However, most existing ONNs are designed for single-object classification, and their performance deteriorates significantly in the presence of multiple objects, limiting their practical applications in multi-object recognition tasks.
In a recent paper published in Light: Science & Applications, a research team led by Professor Nan Zhang from the School of Optics and Photonics at Beijing Institute of Technology developed an anti-interference diffractive deep neural network. This network can accurately and robustly recognize target objects in multi-object scenarios, even under intra-class, inter-class, and dynamic interference.
By employing different deep-learning-based training strategies for targets and interference, the system uses two transmissive diffractive layers to form a physical network that maps the spatial information of targets all-optically into the output light's power spectrum, while dispersing interference as background noise. Validated in the Terahertz band, the designed metasurface can recognize unknown 6-class handwritten-digit under dynamic scenarios involving 40 categories of interference, achieving an experimental testing accuracy of 86.7%.
The researchers highlight three key contributions of their novel network:
(1) High recognition accuracy across diverse scenarios – The method has been tested in various complex settings, such as recognizing one target amid two or three dynamic interferences, and across public datasets including MNIST, Fashion-MNIST, and Quick, Draw!, demonstrating strong generalization capability.
(2) Lightweight metasurface design with strong scalability – The proposed metasurface framework exhibits excellent scalability and can be physically scaled to near-infrared and visible wavelengths, allowing device sizes to be reduced to sub-millimeter scales. This offers a viable path toward highly integrated, low-power optical sensing for edge deployment.
(3) Potential for complex multi-object recognition – By integrating multidimensional optical multiplexing technologies and shift-invariant modules such as optical convolution operators, the system can support recognition of spatially overlapping objects and enable dynamic multi-object recognition.
“This work can significantly advance the practical application of ONNs in target recognition and pave the way for the development of real-time, high-throughput, low-power all-optical computing systems, which are expected to be applied to autonomous driving perception, precision medical diagnosis, and intelligent security monitoring.” The team concludes.
Journal
Light Science & Applications
Article Title
Anti-interference diffractive deep neural networks for multi-object recognition