电子测量与仪器学报

史雨馨,朱继杰,凌志刚.基于特征增强 YOLOv4 的无人机检测算法研究[J].电子测量与仪器学报,2022,36(7):16-23

基于特征增强 YOLOv4 的无人机检测算法研究

Research on UAV detection method based onfeature enhanced YOLOv4 algorithm

DOI：

中文关键词: 卷积神经网络深度学习 YOLOv4 无人机检测特征增强模块

英文关键词:convolutional neural network deep learning YOLOv4 UAV detection feature enhancement module

基金项目:国家自然科学基金（61971183）、湖南省自然科学基金（2021JJ30142）项目资助

作者	单位
史雨馨	1.湖南大学电气与信息工程学院
朱继杰	1.湖南大学电气与信息工程学院
凌志刚	1.湖南大学电气与信息工程学院

Author	Institution
Shi Yuxin	1.College of Electrical and Information Engineering, Hunan University
Zhu Jijie	1.College of Electrical and Information Engineering, Hunan University
Ling Zhigang	1.College of Electrical and Information Engineering, Hunan University

摘要点击次数: 696

全文下载次数: 1320

中文摘要:

现有基于深度学习的目标检测方法在面对空中消费级无人机时,存在鲁棒性差、准确率不足等问题。对此,提出一种基于特征增强的 YOLOv4 目标检测方法—FEM-YOLOv4。首先,针对无人机低、小、慢等特点,改进骨干网络,降低下采样倍数,充分利用包含细粒度信息的浅层特征;其次,加入特征增强模块(feature enhancement module),通过使用不同空洞率的多分支卷积层结构,综合不同深度的语义信息和空间信息,增强小尺度无人机的细节语义特征;另外,利用多尺度融合的特征金字塔结构, 突出特征图包含的细节信息和语义信息,提升模型对无人机目标的预测能力;最后,采用 K-means++算法对无人机目标候选框的尺寸进行聚类分析。与 6 种目标检算法进行对比,实验结果表明,FEM-YOLOv4 算法的 mAP 和 Recall 分别达到 89. 48%、 97. 4%,优于其他算法,且平均检测速度为 0. 042 s。

英文摘要:

Consumer-level UAVs have small scale, low fly speed and height, existing deep learning methods hardly achieve high detection accuracy and good robustness on detecting UAVs. In order to address this problem, this paper develops an improved YOLOv4 algorithm with feature enhanced module named as FEM-YOLOv4 for UAVs detection. Firstly, according to the characteristics of UAVs, this paper reduces the subsampling multiple of CSPDarkNet to improve the backbone network and make full use of shallow features containing detailed information. Secondly, this paper introduces the feature enhancement module to replace the SPP module. The feature enhancement module includes multiple branches and dilated convolution, and it obtains different levels of semantic information, which is beneficial to enhance the detailed semantic features and the detection capabilities of the network. Thirdly, delete the PAN module to improve the feature pyramid, and compress the depth of each detection layer to highlight the detailed and semantic information of the feature maps. Finally, the anchor box is initialized by the K-means++ algorithm to make the model more suitable for predicting the UAV targets. Compared with the six target detection algorithms, the experimental results show that the mAP and Recall of FEM-YOLOv4 algorithm reach 89. 48% and 97. 4% respectively, which are superior to other algorithms, and the average detection speed is 0. 042 s.

查看全文查看/发表评论下载PDF阅读器