High accuracy recognition of the orbital angular momentum (OAM) of light based on petal interference patterns is demonstrated using a convolutional neural network (CNN) approach with an improved Alexnet structure. A type of hybrid beam carrying OAM is utilized to provide more controllable degrees of freedom to recognize the OAM of light. The relationship between the training sample resolution (or the number associated with the accuracy) and the training time of the model, is presented. The recognition accuracy is closely related with the quantum number l of OAM, the angular ratio n of the spire phase over the hybrid phase in one modulation period, and the propagation distance z. Our studies show that when l ranges from 1 to 10, and n varies from 0.02 to 0.99, the recognition accuracy rate of OAM is nearly 100%. The minimum interval of n recognized at the OAM modes decreases to 0.01, which shows the super-high bandwidth of the generation and detection of OAM modes. Such results suggest great potential for the next generation of CNN-based OAM optical communication applications.

中文翻译：

---

神经网络方法实现的光学涡旋模式的宽带宽和高效识别

使用具有改进的Alexnet结构的卷积神经网络（CNN）方法，证明了基于花瓣干涉图案的光的轨道角动量（OAM）的高精度识别。一种类型的携带OAM的混合光束用于提供更多可控制的自由度来识别光的OAM。给出了训练样本分辨率（或与准确性相关的数字）与模型训练时间之间的关系。识别精度与OAM的量子数l，一个调制周期内尖峰相在混合相上的角比n以及传播距离z密切相关。我们的研究表明，当l介于1到10之间时，n在0.02到0.99之间变化，OAM的识别准确率接近100％。在OAM模式下识别的n的最小间隔减小到0.01，这表明OAM模式的生成和检测具有超高带宽。这样的结果表明了下一代基于CNN的OAM光通信应用的巨大潜力。