Multi-Component Feature Extraction for Few-Sample Automatic Modulation Classification

With the rapid development of deep learning (DL), Automatic Modulation Classification (AMC) has also taken a huge leap forward. The DL-based AMC methods are able to achieve high accuracy through training massive labeled samples. However, these DL-based AMC methods would deteriorate dramatically with insufficient samples. Modulation recognition under few sample condition gradually become an urgent problem. To address this problem, we propose a novel learning framework for few-sample AMC, which is termed Multi-Component Extraction Network (MCENet) and can effectively extract potentially and easily distinguishable features. Experimental results on the public available dataset RadioML2016.10a show that the proposed MCENet outperforms other contrastive few-sample AMC methods and achieves better results.