Deep Learning-Based OSNR Estimation from Constellation Diagrams of DP m-PSK and DP m-QAM in Flexible Coherent Optical Receivers

Accurate OSNR estimation is essential for maintaining signal quality in coherent optical communication systems. This work proposes a deep learning-based method for OSNR estimation using constellation diagrams as input. A dataset of over 19,000 images was generated through simulations with various modulation formats. We evaluated 15 CNN architectures, including MobileNetV3, ConvNeXt, DenseNet, and EfficientNet. ConvNeXtBase achieved the best results, with a MAE below 0.43 dB and R2 above 0.98. The results demonstrate the effectiveness of computer vision models for accurate and non-intrusive OSNR prediction.