Automatic Modulation Recognition (AMR) is a technique for automatically determining the modulation type of signals, which is essential to intelligent wireless communication. Recently, deep learning (DL) techniques hav...
详细信息
Automatic Modulation Recognition (AMR) is a technique for automatically determining the modulation type of signals, which is essential to intelligent wireless communication. Recently, deep learning (DL) techniques have significantly facilitated the advancement of AMR methods. While DL models have made remarkable improvements on recognition accuracy for AMR, their implementations on embedded and edge devices are limited by their high computational complexity and huge number of parameters. Lightweight DL-based AMR models are therefore being explored gradually, but the lightweight is often obtained at the expense of accuracy. In this paper, we present a cross-scale feature fusion enhanced multi-level recurrent convolutionalneuralnetwork, and demonstrate its advantages in both saving network parameters and improving recognition accuracy. A simplified recurrent convolutional layer (SRCL) is proposed to extract spatial context information without increasing network parameters. A cross-scale attention enhanced feature fusion layer (AEFF) is developed to emphasize the effective learning of both important local details and global key connections. A multi-level lightweight feature extractor is designed which utilizes multiple SRCLs in parallel to extract features at various levels, and utilizes the AEFF to integrate the extracted multi-scale features with an emphasis of local and global significance. Taking unpreprocessed in-phase and quadrature components of communication signals as input, the proposed model achieves improved recognition accuracy using reduced number of parameters on the benchmark datasets RadioML 2016.10a, 2016.10b and 2016.04c compared with several state-of-the-art DL models for AMR.
暂无评论