基于深度学习的农田害虫目标检测算法的研究与应用  

Research and Application of Farmland Pest Target Detection Algorithm Based on Deep Learning

作　　者：郑睿智 盛猛猛 蔡鑫浩 姚亚洲 唐振民[1] 王琼[1] ZHENG Ruizhi;SHENG Mengmeng;CAI Xinhao;YAO Yazhou;TANG Zhenmin;WANG Qiong(College of Computer science and Engineering,Nanjing University of Science and Technology,Nanjing,Jiangsu 210094,China)

机构地区：[1]南京理工大学计算机科学与工程学院,江苏南京210094

出　　处：《数学建模及其应用》2022年第4期55-61,共7页Mathematical Modeling and Its Applications

基　　金：国家自然科学基金(62102182,61976116);江苏省自然科学基金(BK20210327,BK20220934);中央高校基础研究基金(30920021135)。

摘　　要：为了实现自动化检测农田害虫并提高检测效率,提出了一种基于深度学习的农田害虫目标检测算法.针对数据集存在的长尾分布和多尺度问题,采用一系列在线和离线数据增强、多尺度输入等方法进行预处理和训练,并集成SWA和多模型融合等后处理方法微调模型.最后,在测试中模型AP 50数值能够达到0.898,召回率为0.996.In order to realize automatic detection of farmland pests and improve detection efficiency,a target detection algorithm for farmland pests based on deep learning is proposed.Aiming at the long-tailed distribution and multi-scale problems of the data set,methods such as Copy and Paste,Mixup offline data enhancement,and multi-scale input are used for processing.Finally,the SWA and multi-model fusion methods are used to output the final results.

关 键 词：深度学习 农田害虫 目标检测 数据增强 

分 类 号：O29[理学—应用数学]