Using Time Series Encoders with Image Classifiers to Perform Modulation Recognition

This paper discusses an approach to modulation recognition (classification) of RF signals based on time series encoders in conjunction with image classifiers. The encoders used include Gramian angular fields, Markov transition fields, and recurrence plots. Complex signals are converted into color images using a variety of techniques that handle the real and imaginary components. These color images are used to train resnet34 neural networks to perform classification. Testing is performed on clean signals as well as signals that are impaired with white Gaussian noise. Various amounts of training data are also investigated. Results look promising for select systems under favorable noise and training scenarios.