Attention-guided complex denoising network for automatic modulation recognition

Abstract

With the rapid progress in artificial intelligence technology in recent years, deep learning has gradually become the main method in the field of radar signal automatic modulation recognition (AMR). Under harsh condition of lower signal–noise ratio (SNR), extracting the useful features of the noisy radar signal in the time or the time–frequency domain is extremely difficult because of serious noise corruption of the clean radar signals. Considering the complex-valued characteristic of radar signal, we propose an attention-guided complex denoising network (ACDNet) that consists of denoising and recognition modules. By utilizing the denoising module, we can estimate denoised signal in the noisy background, and thus the more distinctive features of the clean signal are extracted. The recognition module uses the extracted information to complete the modulation recognition task. Furthermore, squeeze-and-excitation blocks (channel attention mechanism) and the network are merged to guide the network to obtain more efficient performance. The experimental results demonstrate that ACDNet has obvious advantages over comparison deep learning algorithms at lower SNRs.