Orbital angular momentum superimposed mode recognition based on multi-label image classification

Orbital angular momentum (OAM) multiplexing technology has great potential in high capacity optical communication. OAM superimposed mode can extend communication channels and thus enhance the capacity, and accurate recognition of multi-OAM superimposed mode at the receiver is very crucial. However, traditional methods are inefficient and complex for the recognition task. Machine learning and deep learning can offer fast, accurate and adaptable recognition, but they also face challenges. At present, the OAM mode recognition mainly focus on single OAM mode and ±l superimposed dual-OAM mode, while few researches on multi-OAM superimposed mode, due to the limitations of single-object image classification techniques and the diversity of features to recognize. To this end, we develop a recognition method combined with multi-label image classification to accurately recognize multi-OAM superimposed mode vortex beams. Firstly, we create datasets of intensity distribution map of three-OAM and four-OAM superimposed mode vortex beams based on numerical simulations and experimental acqusitions. Then we design a progressive channel-spatial attention (PCSA) model, which incorporates a progressive training strategy and two weighted attention modules. For the numerical simulation datasets, our model achieves the highest average recognition accuracy of 94.9$\mathrm{\%}$and 91.2$\mathrm{\%}$ for three-OAM and four-OAM superimposed mode vortex beams with different transmission distances and noise strengths respectively. The highest experimental average recognition accuracy for three-OAM superimposed mode achieves 92.7$\mathrm{\%}$, which agrees with the numerical result very well. Furthermore, our model significantly outperforms in most metrics compared with ConvNeXt, and all experiments are within the affordable range of computational cost.