基于改进YOLOv5s的航拍红外图像目标识别方法

王悠; 韩立祥; 付贵

基于改进YOLOv5s的航拍红外图像目标识别方法

王悠^1,,
韩立祥¹,
付贵^{1, 2, ,}

1.
中国民用航空飞行学院, 四川广汉 618307
2.
四川省通用航空器维修工程技术研究中心, 四川广汉 618307

基金项目:

中央高校基本科研业务费基金项目 J2022-024

四川省通用航空器维修工程技术研究中心资助课题 GAMRC2021ZD01

详细信息

作者简介:
王悠（1978-），女，广西陆川人，高级实验师，硕士生导师，研究方向：计算机技术协同应用研究。Email：yoyo_wang@126.com

通讯作者:
付贵（1990-），男，博士研究生，讲师，主要从事深度学习、图像处理、无人机视觉伺服控制等研究。Email：abyfugui@163.com

中图分类号: TP391.4
计量
- 文章访问数: 115
- HTML全文浏览量: 25
- PDF下载量: 62
出版历程
- 收稿日期: 2023-04-16
- 修回日期: 2023-05-16
- 网络出版日期: 2024-07-24
- 刊出日期: 2024-07-19

Aerial Infrared Image Target Recognition Method Based on Improved YOLOv5s

WANG You^1,,
HAN Lixiang¹,
FU Gui^{1, 2, ,}

1.
Civil Aviation Flight University of China, Guanghan 618307, China
2.
General Aircraft Maintenance Engineering Technology Research Center of Sichuan, Guanghan 618307, China

摘要

摘要:
为了提高无人机在黑暗条件下的识别效率，降低在复杂环境及路况方面存在漏检及延时效果等问题，本文提出了一种改进的YOLOv5s-GN-CB红外图像识别方法，该方法可以提高无人机红外航拍图像对车、人等多类目标识别效率。本文对YOLOv5s的主要改进包括以下3个方面：将Ghost引入到YOLOv5s主干网络中，并将NWD loss损失函数融入至Ghost中；添加注意力机制CA；添加加权双向特征金字塔BiFPN。经验证，改进的YOLOv5s-GN-CB检测模型在InfiRay红外航拍人车检测数据集下目标识别平均精度均值（mAP@0.5）达到95.1%，FPS提高至75.188帧/s。相较于YOLOv5原始模型的平均精度均值和FPS分别提高了4.2%和12.02%。在对同一场景中无人机航拍红外图像目标识别的检测精度有明显提升，延时率有所下降。
- 红外目标检测 /
- 改进YOLOv5s /
- Ghost网络 /
- 注意力机制
Abstract:
To enhance the recognition efficiency of UAVs in dark conditions and reduce missed detections and delays in complex environments and road conditions, this study proposes an improved YOLOv5s-GN-CB infrared image recognition method. This method enhances the efficiency of UAV infrared aerial images for detecting vehicles, people, and other types of targets. The main improvements to YOLOv5s achieved in this study include the following three aspects: 1) introducing the Ghost module into the YOLOv5s backbone network and incorporating NWD loss into Ghost; 2) adding the coordinate attention (CA) mechanism; 3) incorporating a weighted bidirectional feature pyramid network (BiFPN). The improved YOLOv5s-GN-CB detection model achieves an average accuracy of 95.1% (mAP@0.5) on the InfiRay infrared aerial photography man-vehicle detection dataset, with the FPS increased to 75.188 frames per second. Compared with the original YOLOv5 model, the average accuracy and FPS are improved by 4.2% and 12.02%, respectively. In the same scenario, the detection accuracy of UAV aerial photography infrared image target recognition has been significantly improved, and the delay rate has decreased.
- infrared object detection /
- improved YOLOv5s /
- ghost network /
- attention mechanism

HTML全文

图 1 Ghost Bottleneck模块原理图

Figure 1. Module schematic of the Ghost Bottleneck

下载: 全尺寸图片幻灯片

图 2 Ghost module模块原理图

Figure 2. Module schematic of the Ghost module

下载: 全尺寸图片幻灯片

图 3 CA注意力模块

Figure 3. CA attention module

下载: 全尺寸图片幻灯片

图 4 C×H×1和C×1×W的特征图

Figure 4. C×H×1 and C×1×W characteristic diagram

下载: 全尺寸图片幻灯片

图 5 BiFPN注意力模块

Figure 5. BiFPN attention module

下载: 全尺寸图片幻灯片

图 6 改进后YOLOv5s网络结构

Figure 6. Improved YOLOv5s network structure

下载: 全尺寸图片幻灯片

图 7 检测结果示例：两个场景下各改进各算法的检测效果

Figure 7. Example of detection results: The detection effect of each algorithm is improved in two scenarios

下载: 全尺寸图片幻灯片

图 8 改进前后不同模型的参数变化对比

Figure 8. Comparison of parameter variations of different models before and after improvement

下载: 全尺寸图片幻灯片

表 1 消融实验结果对比

Table 1 Comparison of ablation experimental results

Models	mAP@0.5	F1	FPS
YOLOv5s	90.9	87	67.114
YOLOv5s- GC	93.9	89	75.188
YOLOv5s- GN	94.1	89	74.627
YOLOv5s- CB	94.2	90	44.053
YOLOv5s- GN-CB	95.1	91	75.188

下载: 导出CSV

表 2 主流算法对比实验结果

Table 2 Comparative experiments with mainstream algorithms

Models	mAP@0.5/ (%)	FPS/(frame/s)	Weight coefficient/MB
YOLOv5s	90.9	67.114	3.69
YOLOv5s-Ghost	93.9	75.188	7.44
YOLOv5s-GN	94.1	75.188	11.5
YOLOv5s-MobileNetV3	92.0	56.180	7.31
YOLOv5x	97.7	14.164	171
Ours	95.1	75.188	11.6

下载: 导出CSV

参考文献(17)

[1]	张志强, 王萍, 于旭东, 等. 高精度红外热成像测温技术研究[J]. 仪器仪表学报, 2020, 41(5): 10-18. https://www.cnki.com.cn/Article/CJFDTOTAL-YQXB202005002.htm ZHANG Zhiqiang, WANG Ping, YU Xudong, et al. Research on high-precision infrared thermal imaging temperature measurement technology [J]. Chinese Journal of Scientific Instrument, 2020, 41(5): 10-18. https://www.cnki.com.cn/Article/CJFDTOTAL-YQXB202005002.htm
[2]	Redmon J, Farhadi A. YOLOv3: an incremental improvement[EB/OL]. 2018-04-08, [2022-08-12]. https://arxiv.org/abs/1804.02767.
[3]	LIU W, Anguelov D, Erhan D, et al. SSD: single shot MultiBox detector[C]//Computer Vision-ECCV, 2016: 21-37.
[4]	Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[C]// IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016: 779-788.
[5]	ZHAO Z Y, YANG X X, ZHOU Y CH, et al. Real-time detection of particleboard surface defects based on improved YOLOV5 target detection[J]. Scientific Reports, 2021, 11(1): 1-15. DOI: 10.1038/s41598-020-79139-8
[6]	WANG J, CHEN Y, GAO M, et al. Improved YOLOv5 network for real-time multi-scale traffic sign detection[J/OL]. arXiv: 2112.08782, https://arxiv.org/abs/2112.08782
[7]	WEI Yi, LI Xiaofei, LIN Lihui, et al. Causal discovery on discrete data via weighted normalized wasserstein distance[J]. IEEE Transactions on Neural Networks and Learning Systems, 2024, 35(4): 4911-4923. Doi: 10.1109/TNNLS.2022.3213641
[8]	HOU Q, ZHOU D, FENG J. Coordinate attention for efficient mobile network design[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 13713-13722.
[9]	TAN M, PANG R, LE Q V. Efficientdet: scalable and efficient object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 10781-10790.
[10]	LIU S, QI L, QIN H, et al. Path aggregation network for instance segmentation[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 8759-8768.
[11]	Nawaz Marriam, Nazir Tahira, Baili Jamel, et al. CXray-EffDet: chest disease detection and classification from X-ray images using the efficientdet model[J]. Diagnostics, 2023, 13(2): 248-248. DOI: 10.3390/diagnostics13020248
[12]	HAO Hongyun, FANG Peng, DUAN Enze, et al. A dead broiler inspection system for large-scale breeding farms based on deep learning[J]. Agriculture, 2022, 12(8): 1176-1176. DOI: 10.3390/agriculture12081176
[13]	郭吉朋. 基于YOLOv3模型的实时目标检测算法的FPGA实现[D]. 武汉: 华中科技大学, 2020. DOI: 10.27157/d.cnki.ghzku.2020.004780. GUO Jipeng. FPGA implementation of real-time object detection algorithm based on YOLOv3 model[D]. Wuhan: Huazhong University of Science and Technology, 2020. DOI: 10.27157/d.cnki.ghzku.2020.004780.
[14]	Aliyi Salih, Dese Kokeb, Raj Hakkins. Detectio of gastrointestinal tract disorders using deep learning methods from colonoscopy images and videos[J]. Scientific African, 2023, 20: e01628. DOI: 10.1016/j.sciaf.2023.e01628
[15]	王新彦, 易政洋. 基于改进YOLOv5的割草机器人工作环境障碍物检测方法研究[J]. 中国农机化学报, 2023, 44(3): 171-176. DOI: 10.13733/j.jcam.issn.2095-5553.2023.03.024. WANG Xinyan, YI Zhengyang. Research on obstacle detection method of mowing robot working environment based on improved YOLOv5[J]. Journal of Chinese Agricultural Mechannization, 2023, 44(3): 171-176. DOI: 10.13733/j.jcam.issn.2095-5553.2023.03.024.
[16]	徐正军, 张强, 许亮. 一种基于改进YOLOv5s-Ghost网络的交通标志识别方法[J]. 光电子·激光, 2023, 34(1): 52-61. DOI:10.16136/j.joel. 2023.01.0216. XU Zhengjun, ZHANG Qiang, XU Liang. Research on obstacle detection method for working environment of lawn mower robot based on improved YOLOv5[J]. Journal of Optoelectronics Laser, 2023, 34(1): 52-61. DOI:10.16136/j.joel. 2023.01.0216.
[17]	党宏社, 党晨, 张选德. 基于改进YOLOv5s的交通标志识别算法[J]. 实验技术与管理, 2022, 39(9): 97-102. DOI: 10.16791/j.cnki.sjg.2022.09.016. DANG Hongshe, DANG Chen, ZHANG Xuande. Traffic sign recognition algorithm based on improved YOLOv5s[J]. Experimental Technology and Management, 2022, 39(9): 97-102. DOI: 10.16791/j.cnki.sjg.2022.09.016.

施引文献(4)

期刊类型引用(3)

1.	杨旭，李氡，段军勇，姚淼. 基于物探技术的矿区岩土自动化检测数据采集系统设计. 自动化与仪器仪表. 2024(03): 144-148 . 百度学术
2.	戎靖，李入作，王杰，谢欣. 基于卡尔曼滤波算法的电机速度数据自动采集系统设计. 电子设计工程. 2024(22): 63-67 . 百度学术
3.	韩颖. 基于卡尔曼滤波的MEMS陀螺噪声抑制算法设计与实现. 电子测试. 2021(19): 51-53+40 . 百度学术