Neural network classification of structured light in optical turbulence

Abstract

We consider a design of a high-speed wireless optical communication system that involves generating an alphabet using structured light by encoding the transmitted information using Laguerre-Gaussian (LG) beams carrying orbital angular momentum (OAM) [1]. LG beams carrying OAM are chosen for this system due to their resilience when propagated through complex environments. Following propagation through optical turbulence, light intensity patterns are decoded with high accuracy using a convolutional neural network (CNN), commonly used for image classification applications. CNNs only see and learn from the diversity of the received images, they do not recognize how the light intensity distribution has been generated. This motivates us to investigate how the CNN handles and reacts to different forms of structured light that have the same intensity pattern on reception and are transmitted through the same environmental conditions. To compare the effects of structured light on the performance of a CNN we constructed an alphabet using LG beams carrying OAM [2] and 2D projected Far Field Images (FFI) [5]. Under the conditions where there is no induced optical turbulence, the received images have approximately the same intensity pattern, despite differing formation methods. Each form of the structured light is generated using a spatial light modulator (SLM). The slight differences in the generated intensity patterns are a result of the optical artifacts of the SLM and the method used to create the SLM phase screen. The primary comparison metric in this research is the classification accuracy of the CNN when individually trained on images of each type of light. This will provide insight into which form of light is potentially the most resilient to be utilized in receivers supported by CNN. Our experimental set uses a 632.8 nm He-Ne laser and an SLM to generate structured light and propagate two forms of structured light in underwater optical turbulence over a ~2.5 m propagation path. Induced underwater optical turbulence is created using a heater that allows for the control and estimation of turbulence strength. The images are captured by a camera, and the CNN is comprised of 15 layers, reducing computational complexity [4]. The beams are propagated through strong optical turbulence, corresponding to 𝐶𝑛 2 values of approximately 10⁻¹¹ 𝑚^−2/3. Based on preliminary testing, the CNN has been able to accurately learn the resulting variations for all forms of light on as little as 50 images. This result further strengthens the evidence for the resilience of machine learning-based communication systems in harsh environments.