《环境昆虫学报》

作者	单位
杨红飞，杨华，刘付文婷，江楠，邱国葳，巫俊达，徐金柱	1.广东文艺职业学院，广州511400；2. 广东省森林培育与保护利用重点实验室/广东省林业科学研究院，广州 510520；3.广州医科大学生物医学工程学院，广州 511436

摘要点击次数: 6

全文下载次数: 2

中文摘要:

海榄雌瘤斑螟Ptyomaxia syntaractis，红树植物白骨壤Avicennia marina最重要害虫，严重影响白骨壤生长和生态功能的发挥。为高效监测海榄雌瘤斑螟的种群发生动态，实时获得预警信息，本研究通过引入目标检测算法YOLO V5进行深度学习，对监测设备上的海榄雌瘤斑螟进行识别与计数，实时发布种群数量。采用黑光灯诱捕装置获取海榄雌瘤斑螟成虫图像，构建两种不同图像大小的数据集，采用旋转、增噪等方式增强图像数据集；对比了不同训练模型对采集图像的检测性能和不同图像大小对数据集识别结果的影响，用精确率、召回率、F1值、平均精度评估各模型的差异。测试结果表明，模型YOLO V5s对海榄雌瘤斑螟识别的精确率、召回率和F1值分别为96.13%、92.06%和0.93，并且能够很好的识别原始尺寸的图像。基于YOLO V5网络模型设计的海榄雌瘤斑螟识别计数模型识别准确率高，可满足海榄雌瘤斑螟种群监测与预警。

英文摘要:

Ptyomaxia syntaractis, the main pest of the mangrove plants Avicennia marina, affected the growth and ecological function of A. marina seriously. In order to efficiently monitor the population dynamics, obtain the early warning information and publish population numbers in real time, object detection algorithm YOLO V5 was introduced for deep learning to identify and count the moth on the monitoring equipment in this study. Black light trapping devices were used to obtain the adult images of P. syntaractis, and two datasets with different image sizes, enhanced by means of rotation and noise enhancement were constructed. The detection performance of different training models on acquired images and the effect of different image sizes on the recognition results of datasets were compared, and accuracy, recall rate, F1 value and average accuracy were used to evaluate the differences among the models. The 〖LM〗results showed that the accuracy, recall rate and F1 value of YOLO V5s model for the identification P. syntaractis were 96.13%, 92.06% and 0.93 respectively, and the model could well recognize the original size image. The identification and counting model based on YOLO V5 algorithm can be used in the population monitoring for its high recognition accuracy.

查看全文查看/发表评论下载PDF阅读器

关闭