文献类型：期刊文献

中文题名：基于YOLOv5的红外相机野生动物图像识别

英文题名：Recognition of Wild Animals Using Infrared Camera Images Based on YOLOv5

作者：杨铭伦[1,2] 张旭[1,2] 郭颖[1,2] 于新文[1,2] 侯亚男[1,2] 高家军[1,2]

第一作者：杨铭伦

通信作者：Guo, Y[1];Guo, Y[2]

机构：[1]中国林业科学研究院资源信息研究所,北京100091;[2]国家林业和草原局林业遥感与信息技术重点实验室,北京100091

年份：2022

卷号：59

期号：12

起止页码：372-380

中文期刊名：激光与光电子学进展

外文期刊名：Laser & Optoelectronics Progress

收录：CSTPCD;;Scopus;WOS:【ESCI(收录号:WOS:000823188100038)】；北大核心:【北大核心2020】；CSCD:【CSCD2021_2022】；

基金：中国林业科学研究院中央级公益性科研院所基本科研业务费专项资金;海南长臂猿智能感知技术研究与应用示范(CAFYBB2019ZB011)。

语种：中文

中文关键词：机器视觉;野生动物;目标检测;YOLOv5;图像识别;深度学习

外文关键词：machine vision;wild animal;target detection;YOLOv5;image recognition;deep learning

分类号：TP391.4

摘要：为在红外相机等资源受限平台上实时、准确地实现海量野生动物图像自动识别,改善野生动物监测过程中数据传输负载重、时效性低等问题,基于YOLOv5模型,利用5类物种的红外相机图像构建数据集,对YOLOv5s、YOLOv5m、YOLOv5l、YOLOv5x四种网络结构进行训练。通过对比不同网络结构的精度、检测速度、体积,明确最优网络结构;同时分析模型在复杂背景信息干扰下的识别效果,评价YOLOv5在真实野外场景的适用性;并通过与其他同类算法的比较,明确YOLOv5用于野生动物识别的优势。实验结果表明:四种网络结构的识别精度均较高,F1-score和平均精度(mAP)均在90%以上,其中YOLOv5m的综合性能最好;YOLOv5在多种复杂背景信息干扰下识别效果仍较好,能够很好地适应真实野外场景;与其他算法相比,YOLOv5同时具有精度高、鲁棒性强、资源占用低等优势。YOLOv5是一种轻量化的模型且性能优越,为在资源受限的平台上进行野生动物实时识别提供了新的契机。
In this paper, we propose the construction of an extended YOLOv5 model using the infrared camera image datasets of five species to achieve the automatic recognition of massive wild animal images in real-time,accuracy on resource-limited platforms such as infrared cameras. Furthermore, we improve the negative load and low timeliness of data transmission in wildlife monitoring. Here, the dataset constructed is used to train four network structures, namely, YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x. By comparing the accuracy, detection speed, and volume of different network structures, the optimal network structure was determined. Simultaneously,we analyzed the recognition effect of the model under the interference of complex background information to evaluate the applicability of YOLOv5 in real-field scenes. Compared with similar algorithms, the advantages of YOLOv5 for wildlife recognition outweighed others. The experimental results show that the recognition accuracy of the four network structures was high. Moreover, F1-score and average accuracy(mAP) were more than 90%, and the comprehensive performance of YOLOv5m was the best. However, YOLOv5 still has a good recognition effect under the interference of several complex background information and can adapt to real-field scenes. Compared with other algorithms, YOLOv5 has the advantages of high precision, strong robustness, and low resource occupation. It is a lightweight model with superior performance, which provides a new opportunity for real-time wildlife identification on resource-constrained platforms.