Higher-order spatial mode detection leveraging deep learning on random optical patterns

Abstract

The speckle based classification of Laguerre-Gauss (LG) beams using the convolution neural network (CNN) has gained a significant interest in view of their applications towards free space optical communication. For the LG beams with higher azimuthal index or topological charge (TC), the classification becomes extremely difficult due to the presence of similarities or structural correlations between the speckle patterns corresponding to the various orders. Here, we propose two deep learning techniques based on CNN and a Vision Transformer (ViT) for classifying the LG beams with higher TC up to 50 using the speckle patterns generated by scattering these modes through a rough surface. The obtained results show that the ViT based model achieves a better mean classification accuracy of 98.1%, outperforming the CNN based model by approximately 3.3%. Additionally, we validated the robustness of the proposed models by accurately detecting TC over three independent test runs and we are also able to determine the sign of TC as well. This classification was performed using a small patch (224 $$\times $$ 224) of speckle patterns, approximately 1% of the recorded speckle pattern. The robustness of the proposed method for high topological charges after perturbing using the scattering environments enables more reliable, high capacity optical systems for real world use.