改进YOLOv5的闪电哨声波轻量化自动检测模型

Lightweight Automatic Detection Model for Lightning Whistle Waves Based on Improved YOLOv5

提出一种改进YOLOv5(You-Only-Look-Once version 5)检测模型YOLOv5-Upgraded.为了更快定位真实边框,该模型将损失函数CIoU (Complete IoU)替换为SIoU(Scylla IoU);同时为了避免网络训练过程中梯度消失、梯度爆炸以及神经元坏死等现象,将激活函数SiLU(Sigmoid-weighted Linear Unit)替换为具有更好梯度流的Mish;在主干网络中插入注意力(Coordinate Attention, CA)机制,帮助模型更精准地识别闪电哨声波,大大降低了漏检率.基于张衡一号感应磁力仪(Search Coil Magnetometer, SCM)数据,以2.4 s时间窗口截取数据,经带通滤波、短时傅里叶变换得到1126张时频图数据集,再经图像增强操作扩充至7882张,其中7091张作为训练集, 791张作为测试集.实验结果表明,基于改进YOLOv5的模型平均精度均值为99.09%,召回率为96.20%,与YOLOv5s相比,分别提升了2.75%和5.07%,与基于时频图的YOLOv3模型相比,平均精度均值和召回率则分别提高了5.89%和9.62%.基于智能语音的LSTM (Long Short Term Memory Networks)闪电哨声波识别模型大小为82.89 MB, YOLOv5-Upgraded仅为13.78 MB,约节省83.38%的内存资源.研究表明改进后的轻量化模型大大降低了闪电哨声波的漏检现象,在测试集中取得了较好结果,并且其轻量化特征易于部署到卫星设备,极大地提高了星载识别的可能性.

This project proposes an improved YOLOv5 detection algorithm YOLOv5 Upgraded.To address this issue,the study proposes an improved YOLOv5 detection algorithm called YOLOv5-Up-graded.The model takes into account the vector angle between the predicted edge and the real edge,The model replaces the loss function CIoU(Complete IoU)with SIoU(Scylla IoU);at the same time,in or-der to avoid phenomena such as gradient disappearance,gradient explosion,and neuron necrosis during network training,the activation function SiLU(Sigmoid-weighted Linear Unit)is replaced with Mish with better gradient flow;The CA attention mechanism is inserted into the backbone network to help the model identify the Lightning whistler waves more accurately and greatly reduce the missed detec-tion rate.The study is based on the VLF-band data of CSES Satellite SCM with 2.4 seconds time win-dow to intercept data,and 1126 time-frequency map data sets are obtained by band-pass filtering and short-time Fourier transform,and then expanded to 7882 images by image enhancement operations,of which 7091 are used as training set and 791 are used as test set.Experimentally,the average mean accu-racy(mAP)of the improved YOLOv5-based model is 99.09%and the Recall is 96.20%,which are im-proved by 2.75%and 5.07%compared with the plain YOLOv5s,and 5.89%and 9.62%compared with the time-frequency map-based YOLOv3 model.The size of LSTM based on the speech processing tech-nology lightning whistler waves recognition model is 82.89MB,while the YOLOv5-Upgraded model is on-ly 13.78 MB,saving about 83.38%of memory resources.It is shown that the model greatly reduces the leakage problem of Lightning whistler waves,achieves better results in test set,and its lightweight fea-tures are easy to deploy to satellite devices,which greatly improves the possibility of satellite recogni-tion.