采用混合域注意力机制的无人机识别方法

Drone Identification Method Based on Mixed Domain Attention Mechanism

针对在城市公园、广场和大型游乐场等公共环境中,雷达和无线电识别无人机易受到电子干扰、图像识别无人机易受到光线和遮挡物干扰的问题,提出了一种经济便捷、不易受到干扰的运用声音和采用通道空间混合域注意力机制多尺度分组卷积网络(ECSANet)的无人机识别方法。首先,建立民用的9大类无人机声音数据集,提取数据集的对数梅尔谱图及其动态特征;其次,为了网络参数量少,避免过拟合,设计了基于分组卷积、通道混洗和残差结构的通道混洗多尺度分组卷积网络(MSSGNet);然后,为了能更多、更有效地提取无人机声音特征,设计了通道空间混合域注意力机制模块(ECSA);最后,将ECSA模块插入MSSGNet网络构成改进的通道空间混合域注意力机制的多尺度分组卷积网络(ECSANet),形成新型声音识别无人机的方法。运用设计的ECSANet网络对自建的民用无人机声音数据集和Urbansound8K环境声音数据集进行了声音识别,识别结果表明:与ResNet18、ResNet34、ResNeXt18和MobileNetV2等基准网络相比,MSSGNet网络参数更少,识别准确率更高,达到了95.1%;ECSA模块可以插入多种网络,在不增加很多参数的情况下令网络模型的识别准确率获得提升,在无人机等声音分类任务上具有很好的效果;与MSSGNet网络相比,改进的ECSANet网络识别准确率能达到95.9%,提高了0.8%,表明了该网络在识别小样本无人机方面的优越性和可行性。

An economical,convenient and undisturbed drone detection method using sound and multiscale group convolution network with attention mechanism in mixed domain of channel space(ECSANet)is proposed in the context of susceptibility to electronic interference in identification of drones by radar and radio,and the interference of light and obstruction in identification of drones by images in public environments such as urban parks,squares and large amusement parks.Firstly,nine kinds of sound dataset of civil drones are established,and their logarithmic Mel spectra and dynamic characteristics are extracted.Secondly,based on packet convolution,channel shuffling and residual structure,a multi-scale group convolution network with channel shuffle(MSSGNet)is designed to reduce the network parameters and avoid over fitting.Then,the efficient channel and spatial attention(ECSA)is designed to extract more and more effective features of drone sounds.Finally,the ECSA is inserted into the MSSGNet to form an improved multiscale group convolution network with attention mechanism in mixed domain of channel space(ECSANet),offering a new method for sound recognition of drones.The designed ECSANet is used to identify the self-built civil drone sound dataset and environmental sound dataset urbansound8k.The results reveal that when compared with benchmark networks such as ResNet18,ResNet34,ResNeXt18,and MobileNetV2,the MSSGNet has fewer network parameters but a higher identification accuracy(up to 95.1%).The ECSA can be inserted into a variety of networks to improve identification accuracy of network models without introducing too many parameters,and it works well for sound classification tasks like drones.As compared with the MSSGNet,the improved ECSANet has an identification accuracy of 95.9%,an increase of 0.8 percent,demonstrating the superiority and feasibility in identifying a small sample of drones.