Deep Learning-Based Modulation Recognition for Imbalanced Classification
Runqing Luo, Jian Sun, Yanan Guo · 2023
Automatic modulation recognition (AMR) is a crucial task in intelligent receiver systems, with significant applications in both civilian and military scenarios. Deep learning-based modulation recognition algorithms have achieved high recognition accuracy and good robustness. However, the performance of DL-AMR heavily relies on a large amount of training data. When the training dataset has a limited number of samples or an imbalanced distribution across different classes, the accuracy of automatic modulation recognition for minority class samples may significantly decrease. Therefore, a CNN model with a focal loss as the output layer's loss function is proposed to address the issue of decreased recognition accuracy for minority classes in the deep learning model when faced with imbalanced training data. Simulation results demonstrate that this model improves the recognition accuracy of classes with a limited number of samples.