基于深度学习的海上目标一维序列信号目标检测方法

当前海面目标检测方法多基于统计理论,检测性能受背景统计特性假设的影响,本文从信号预测和特征分类两个角度,分别采用长短时记忆网络(LSTM)和卷积神经网络(CNN)对信号时间序列幅度信息进行处理,用于海上目标一维序列雷达信号检测,该方法不需事先假设背景统计特性,泛化能力更强。基于LSTM序列预测的目标检测方法通过用海杂波信号幅度时间序列对网络进行训练,再用训练后的网络对后续序列进行预测,并与后续实测信号进行比较,实现目标检测。基于CNN序列分类的目标检测方法中采用截取的海杂波信号和目标信号幅度序列作为数据集样本,对一维卷积核CNN进行训练,使其具有识别目标杂波信号特征能力,从而实现目标检测。最后,采用IPIX和CSIR实测海杂波数据对两种方法进行验证,结果表明两种方法均可实现一维序列信号中海面目标的检测,但LSTM预测方法对于长序列检测的实时性有待于进一步提高;CNN分类方法可实现实时检测,但仅利用信号幅度信息,检测性能仍需进一步提升。

One-dimensional Sequence Signal Detection Method for Marine Target Based on Deep Learning

In this paper, the feature generalization learning ability of deep learning is used to process the signal time series amplitude information. From the perspectives of signal prediction and feature classification, respectively, long short-term memory networks (LSTM) and convolutional neural networks (CNN) are used for the detection of target’s one-dimensional sequence radar signal. The target detection method based on LSTM sequence prediction uses the sea clutter signal amplitude time series to train the network, and then uses the trained network to predict subsequent sequences, and compares it with subsequently measured signals to achieve target detection. In the target detection method based on CNN sequence classification, the intercepted sea clutter signal and the target signal amplitude sequence are used as dataset samples to train the one-dimensional convolution kernel CNN so that it can identify the target clutter signal feature, thereby achieving the target detection. Finally, the two methods were verified using IPIX and CSIR measured sea clutter data. The results show that both methods can detect sea-surface targets in one-dimensional sequence signals, but the real-time performance of LSTM prediction methods for long sequence detection needs to be further improved. CNN classification method can realize real-time detection, but using only signal amplitude information, the detection performance still needs to be further improved.