基于深度学习的白纹伊蚊卵粒识别模型初步研究
PRELIMINARY STUDY ON EGG RECOGNITION MODEL OF AEDES ALBOPICTUS BASED ON DEEP LEARNING
朱敏慧 1董琳娟 1王墩家 1蔡逸舟 1张兆文 1刘曜 2何世鹏 1周毅彬1
作者信息
- 1. 上海市闵行区疾病预防控制中心, 上海 201100
- 2. 上海市疾病预防控制中心, 上海 200336
- 折叠
摘要
目的 构建白纹伊蚊Aedes albopictus卵粒图片数据库,并基于深度学习和瓦片重叠法建立白纹伊蚊卵粒自动识别和计数模型,为伊蚊监测和防治提供技术方法.方法 在上海市 3 个区收集野外和实验室品系的白纹伊蚊卵粒图片 449 张,使用Python环境labelimg库进行人工标定蚊卵,采用快速区域卷积神经网络(faster regional convolution neural network,Faster R-CNN)模型建立白纹伊蚊卵粒识别模型,并使用瓦片重叠法智能识别和计数.使用精确率、召回率和调和平均值(F-measure)进行模型效果评价.结果 经过15 次模型迭代训练,损失值随着训练次数增加逐渐下降至 0.000119,平均精度均值(Mean Aaccuracy,mAP)从 0.968 增加至 0.980,最终模型精确率为 0.90,召回率为 0.97,调和平均值为 0.93.结论 本模型具有较高的白纹伊蚊卵粒识别能力,初步达到了辅助开展诱蚊诱卵器监测中卵粒鉴别和计数的功能.在不断的优化模型、细化分类识别能力后可成为简便高效的监测辅助工具.
Abstract
Objective In this study,a pictorial database of Aedes albopictus eggs was built to establish a model for the automatic recognition and counting of eggs of this mosquito species.Methods In total,449 images of Ae.albopictus eggs from field strains in three districts of Shanghai and laboratory strains were collected.The eggs were manually calibrated using the Python environment labeling library,and the faster region-based convolution neural network(Faster R-CNN)was used to train the model with the tile overlap method.Results The results of model validation evaluation using accuracy,recall,and F-measure showed that after 15 training sessions,the loss gradually decreased with increasing training frequency and ultimately decreased to 0.000119.The mean accuracy(mAP)increased from 0.968 to 0.980 with increasing training frequency.The final model had an accuracy of up to 0.90,a recall rate of 0.97,and an F-measure of 0.93.Conclusion The established model achieved its function of assisting in egg identification and counting.With further optimization of the model and refining of its classification and recognition capabilities,it will serve as a simple and efficient auxiliary monitoring tool.
关键词
深度学习/诱蚊诱卵器/白纹伊蚊/卵粒Key words
Deep learning/Mosq-ovitrap/Aedes albopictus/Eggs引用本文复制引用
基金项目
上海市闵行区自然科学基金研究项目(2023MHZ002)
上海市闵行区公共卫生重点学科建设项目(MGWXK2023-09)
出版年
2024
国家科技期刊平台
NSTL
万方数据