Automatic Modulation Recognition Based on a Dual-Branch Architecture and SNR-Aware Feature Adjustment

Automatic Modulation Recognition (AMR) in noisy environments, particularly at low Signal-to-Noise Ratios (SNRs), is prone to substantial performance degradation. In this paper, we propose a novel framework for robust low-SNR modulation recognition, featuring a dual-branch architecture that concurrently leverages two complementary signal representations: raw I/Q samples, and Amplitude-Phase Continuous Spectrum (APCS). An SNR-FiLM conditioning layer is incorporated to adaptively adjust the network's activations based on SNR levels. Evaluation on the RadioML2016.10a dataset demonstrates an overall accuracy improvement from 61.36% in the baseline model to 67.60%. In low-SNR scenarios, notable gains are achieved, with accuracy at -6 dB and -4 dB increasing from 52.45% and 67.00% to 70.09% and 81.68%, respectively.