A Novel Lightweight Automatic Modulation Classification Scheme Based on Inverted Residuals

Automatic modulation classification is an indispensable part of present and future wireless communication systems. The deep learning helps automatic modulation classification realize superior performance. However, most of the DL-based schemes have a large model scale and their computational complexity is high, which leads to the difficulty in the applications. In order to overcome this challenge, motivated by the lightweight models in computer vision, a novel lightweight AMC schemes based on inverted redisual structure and linear bottleneck is proposed in this paper. The inverted residual structure is employed to extract refined features under the condition of low computational cost. Linear bottleneck avoids the features loss of activation function. Numerous simulation results prove that the proposed lightweight AMC scheme can vastly decrease the computational cost comparing to the benchmark schemes. Additionally, the classification accuracy is guaranteed by using our proposed scheme.