摘要
针对现有基于深度学习理论的信号智能检测方法大多只能对单信号或时频域不重叠的信号进行检测,本文提出了一种基于掩膜区域卷积神经网络(Mask R-CNN)与Criminisi算法的时频重叠多信号智能检测新方法。首先将一维时域信号通过时频变换得到二维时频图像。然后针对时频图中多信号重叠部分像素位置信息缺失这一问题,提出了利用Criminisi算法对信号重叠部分像素位置信息进行恢复。最后,基于缺失信息恢复后的图像使用Mask R-CNN进行训练,再用训练后的网络对未知信号进行检测。实验结果表明,该方法在信噪比(SNR)为-3 dB时,时频域重叠信号的平均检测率达92%,相比基于卷积神经网络的信号检测方法,在SNR大于-3 dB时检测率平均提高20%以上。
In view of the existing intelligent signal detection methods based on deep learning theory,most of them can only detect single signal or signals which don’t overlap in time-frequency domains.This paper proposes a new intelligent detection method based on Mask R-CNN and Criminisi algorithm for time-frequency overlapping multi-signals.First,the signal in the time domain is transformed into a time-frequency image.Then,to solve the problem of missing pixels’position information in the overlapping part of multiple signals in the time-frequency domain,the Criminisi algorithm to repair and fill the information is applied.Finally,Mask R-CNN is used for training the restored image,and used for detecting the unknown signals.Experimental results show that when the SNR is 0 dB,the average detection rate of overlapping signals in the time-frequency domain reaches 99%.Compared with the method based on convolutional neural network,the average detection rate is increased by more than 20%.
作者
李杰
孙闽红
仇兆炀
LI Jie;SUN Minhong;QIU Zhaoyang(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou,Zhejiang 310018,China)
出处
《信号处理》
CSCD
北大核心
2021年第5期878-884,共7页
Journal of Signal Processing
