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.
Lightweight Automatic Detection Model for Lightning Whistle Waves Based on Improved YOLOv5
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.
国家科技期刊平台
NSTL
万方数据
