摘要
为了解决传统的短波信号增强方法存在的拟合能力差和非线性能力差的问题,将传统滤波方法与深度学习的方法相结合。为保留短波信号中更多细粒度特征和增加网络复杂度,在语音增强生成对抗网络架构中加入残差网络,提出基于语音增强生成对抗网络的短波信号增强算法;将维纳滤波方法与短波信号增强生成对抗网络算法相结合,在波形级别上进行端到端的模型训练,实现了将生成对抗网络的衍生算法用于短波信号增强。实验结果表明,增强后的信号样本与带噪信号样本相比,信噪比和清晰度等指标显著提升,证明了所提方法的可行性和有效性。同时,该方法比传统滤波方法适用场景范围大,可作为信号处理的前端环节,为非合作方短波信号研究提供可靠的参考方法。
In order to retain more fine-grained features in short-wave signals and increase the network complexity,the residual network is added into the architecture of speech enhancement generative adversarial network(SEGAN), and a shortwave signal enhancement algorithm based on SEGAN is proposed. The Wiener filtering method is combined with the shortwave signal enhancement generative adversarial network(SSEGAN) algorithm to carry out end-to-end model training at the waveform grade, so that the derived algorithm of generative adversarial network can be used for the short-wave signal enhancement. The experimental results show that the signal-to-noise ratio(SNR) and clarity of the signal samples after enhancement processing are significantly improved in comparison with those of the signal samples with noise,which proves the feasibility and effectiveness of the proposed method. In addition,the range of applicable scenes of the proposed method is wider than that of the traditional methods. Therefore,it can be used for the front end of signal processing,and can provide a reliable reference for the research of non-partner short-wave signals.
作者
张睿琦
李迟生
陈颖
ZHANG Ruiqi;LI Chisheng;CHEN Ying(School of Information Engineering,Nanchang University,Nanchang 330031,China)
出处
《现代电子技术》
2021年第17期56-60,共5页
Modern Electronics Technique
基金
国家自然科学基金项目(61661030)。
关键词
短波信号增强
生成对抗网络
残差网络
深度学习
模型训练
特征提取
信噪比
shortwave signal enhancement
generative adversarial network
residual network
deep learning
model training
feature extraction
SNR