Implementing Digital Recognition Based on Transmissive Optical Neural Networks

Optical Neural Networks (ONNs) have attracted great attention from both academia and industry due to their distinct advantages in power consumption, parallelism, and computational speed. This paper presents a novel ONN system that not only reduces the distances between system components but also enhances network accuracy. By employing fully transmissive optical devices, such as transmissive spatial light modulators, the system utilizes a single optical axis for light propagation. A micro-lens array (MLA) is used to enlarge the input optical field, where the relation between the number of micro-lenses in this MLA and the accuracy of this ONN have been investigated. Moreover, the number of micro-lenses has been optimized according above relationship. Meanwhile, the relation between the important parameters, such as the position and the size, of a detection unit and the accuracy of this ONN has been studied. Using the MNIST handwritten digit dataset as a test case, this ONN achieved an accuracy rate to 96.1%. Finally, this architecture is validated by experiment, which show a great accuracy for digital recognition.