改进的基于端到端学习的蜻蜓目昆虫识别算法  

Improved End-to-End Learning Odonata Recognition Algorithm

作　　者：彭明杰 唐万梅[1] 皮家甜 于昕 吴至友[3] 黄芳婷 PENG Mingjie;TANG Wanmei;PI Jiatian;YU Xin;WU Zhiyou;HUANG Fangting(College of Computer and Information Sciences,Chongqing Normal University;Chongqing Center of Engineering Technology Research on Digital Agriculture Service,Chongqing Normal Universitty;Chongqing Key Laboratory of Intelligent Finance and Big Data Analysis,Chongqing Normal Universitty;College of Life Sciences,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331 [2]重庆师范大学重庆市数字农业服务工程技术研究中心,重庆401331 [3]重庆师范大学智慧金融与大数据分析重庆市重点实验室,重庆401331 [4]重庆师范大学生命科学学院,重庆401331

出　　处：《重庆师范大学学报（自然科学版）》2021年第3期94-99,F0003,F0002,共8页Journal of Chongqing Normal University:Natural Science

基　　金：科学技术部科技基础性工作专项重点项目(No.2015FY210300);重庆市教育委员会科技项目青年项目(No.KJQN201800521);重庆市基础研究与前沿探索项目(No.cstc2018jcyjAX0470);重庆市研究生科研创新项目(No.CYS20247)。

摘　　要：【目的】传统的蜻蜓目(Odonata)昆虫的鉴别算法是在标本图片上进行人工的特征提取来训练分类器,此类方法所用的数据集包含的蜻蜓目昆虫种类和数量均较少,并且样本背景单一,导致识别率低且只能识别标本图片。针对这些问题制作了一个同时包含蜻蜓目昆虫生态图片和标本图片的数据集,提出一种基于深度学习的蜻蜓目昆虫的识别算法。【方法】采用具备端到端学习优势的网络框架,在上述数据集上,使用在ImageNet上迁移学习的ResNet50进行特征提取,使用新的区域建议网络Imp-RPN进行特征筛选,再使用改进的损失函数来解决样本分布不均的问题。【结果】所提出的识别算法在精确到种的46类分类任务中识别率达到了96.2%,在精确到种并包含性别信息的90类分类任务中识别率则达到了88.7%。【结论】端到端的深度学习网络框架免除了人工标注样本的时间成本,提高了识别准确率,更能满足物种鉴别任务的需求。[Purposes]The traditional Odonata insect identification algorithm is to perform artificial feature extraction on specimen pictures to train the classifier.The data set used by this method contains fewer types and numbers of Odonata insects,and the sample background is single,resulting in low recognition rate and can only identify the specimen pictures.In response to these problems,a data set containing both ecological pictures and specimen pictures of Odonata insects was produced,and a recognition algorithm based on deep learning was proposed.[Methods]Using a network framework with end-to-end learning advantages,on the above data set,use ResNet50 migrated and learned on ImageNet for feature extraction,use the new region suggestion network ImpRPN for feature selection,and then use an improved loss function to solve the problem of uneven sample distribution.[Findings]Experimental results show that the recognition rate of the recognition algorithm reaches 96.2%in 46 classification tasks that are accurate to species,and 88.7%in 90 classification tasks that are accurate to species and contain gender information.[Conclusions]The end-to-end deep learning network framework eliminates the time cost of manually labeling samples,improves the recognition accuracy,and can better meet the needs of species identification tasks.

关 键 词：蜻蜓目识别 细粒度识别 深度学习 端到端学习 NTS-Net 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]