Evaluating the Efficacy of Pre-Trained ImageNet Weights for Constellation Diagram Classification Using Deep Learning Models

Constellation diagram classification is crucial for selecting the appropriate demodulation method in communication systems, as it allows for accurate interpretation of modulated signals. In this context, it is essential to investigate the impact of using pretrained ImageNet weights in classical CNN models for this task. This paper explores the effect of ImageNet pretrained weights on the classification of simulated constellation diagrams, including QAM-16, QAM-32, and QAM-64 modulation schemes. Three deep learning models, VGG, ResNet, and MobileNet, were evaluated using accuracy as the primary performance metric. The models were trained both with and without ImageNet pretrained weights to compare their effectiveness. The experimental results demonstrated that incorporating ImageNet weights led to significant performance improvements in most models. MobileNet, in particular, achieved the highest accuracy, showcasing its ability to effectively learn from pretrained weights and generalize well to the constellation classification task. ResNet showed moderate improvements, while VGG failed to benefit from the pretrained weights. These findings suggest that smaller, more efficient models like MobileNet are better suited for tasks with limited data. In conclusion, this study validates the effectiveness of deep learning models for constellation diagram classification and highlights the benefits of using ImageNet pretrained weights to enhance performance, especially for models with fewer parameters. Future work could focus on exploring more complex modulation schemes and expanding the range of models tested.