Recognition of high-resolution optical vortex modes with deep residual learning

Optical vortex beams with fractional orbital angular momentum (OAM) can greatly enhance the channel capacity in free-space optical communication. However, high precision measurement of fractional OAM modes is always difficult, especially under the influence of atmospheric turbulence (AT). In this work, we identify the high-resolution OAM modes down to 0.01 using an improved residual neural network (ResNet) architecture based convolutional neural network (CNN). Experimentally, using a single cylindrical lens, the light intensity distribution can be readily converted into a diffraction pattern containing significant features trained into a CNN model. For the fractional OAM modes from 5.0 to 5.9 over a long propagation distance of 1500 m, at 0.1 resolution, our model's predicting accuracy is up to 99.07% under strong AT, Cn2=1×10-15m-2/3. At 0.01 resolution, the accuracy is as high as 86.98% under intermediate AT, Cn2=1×10-16m-2/3, and exceeds 73.78% under strong AT, Cn2=1×10-15m-2/3. So, these results may have great implications in free-space optical communication.