Title: Learning of Time-Frequency Attention Mechanism for Automatic Modulation Recognition

Abstract: Recent learning-based image classification and speech recognition approaches make extensive use of attention mechanisms to achieve state-of-the-art recognition power, which demonstrates the effectiveness of attention mechanisms. Motivated by the fact that the frequency and time information of modulated radio signals are crucial for modulation mode recognition, this paper proposes a time-frequency attention mechanism for a convolutional neural network (CNN)-based modulation recognition framework. The proposed time-frequency attention module is designed to learn which channel, frequency and time information is more meaningful in CNN for modulation recognition. We analyze the effectiveness of the proposed time-frequency attention mechanism and compare the proposed method with two existing learning-based methods. Experiments on an open-source modulation recognition dataset show that the recognition performance of the proposed framework is better than those of the framework without time-frequency attention and existing learning-based methods.