“世界有害实蝇自动识别系统2.0”的设计、构建及应用

An introduction to the design,construction and application of"The Automated Fruit Fly Identification System 2.0(AFIS2.0)"

实蝇作为一类十分重要的经济类昆虫,其快速和准确鉴定一直是困惑各国植检,以及农林等部门的重要难题。“世界有害实蝇自动识别系统2.0”(AFIS2.0),针对实蝇科8属83种,基于深度学习框架,利用Mask-R-CNN模型对图像分割校准,根据Discriminative Deep Metric Learning原理,对预训练的AlexNet模型进行微调提取分割后特征,采用模板匹配法对图像分类鉴别;基于内嵌BLAST+程序及外源BOLD链接进行分子序列比对;依据固定比例权重融合图像及分子识别结果,构建集成镶嵌翅、胸、腹及分子信息的在线实蝇科自动识别系统。其主要包括数据输入、预处理、自动识别、结果显示及物种复核五模块。此外,本文还对“世界有害实蝇自动识别系统2.0”的主界面、功能菜单、主操作流程及一些其它功能进行了介绍,讨论了影响AFIS2.0识别准确度的因素,总结了主要特点,展望了未来应用和发展前景。经检测翅图像最佳识别率(识别结果列表中的Top1物种)达90%,翅、中胸背板和腹部背面图像的Top5平均识别率为94%。初步应用结果表明,可一定程度减少有害实蝇鉴定所需的搜索范围和鉴定时间,部分解决口岸及农林部门有害实蝇物种的自动鉴定和识别问题。

The rapid and accurate identification of fruit flies(Diptera:Tephritidae)with economic importance has always been a problem for agriculture,forestry and quarantine departments in various countries.The Automated Fruit Fly Identification System 2.0(AFIS2.0)was developed by using 83 species of 8genera in Tephritidae.It was based on the deep learning framework,applied the Mask Region Convolutional Neural Network(Mask R-CNN)and discriminative deep metric learning methods for extracting segmented features by fine-tuning the pre-trained AlexNet models,adopted template matching method for image identification,integrated Blast+and BOLD(The Barcode of Life Data System)link for DNA sequence comparison,used specific weighting for the fusion result of image and DNA sequence.An online intelligent identification system for the fruit fly pests was successfully constructed,which integrated a large amount data of wing,thorax,abdomen images and molecular information in this family.It included five modules:image/DNA sequence input,image preprocessing,automatic identifying,result displaying and species interactive checking.Furthermore,this paper also introduces the main interface,function menu,main operation process and some other functions of AFIS2.0,discusses the factors that affect its recognition accuracy,summarizes the main features,and looks into future on its application and development prospects.Finally,at the species level,the best classification accuracy rate for wing images(as the Top 1species in the species list of outcomes)reached 90%,and the average classification accuracy rate for wing,thorax,and abdomen images(as the Top 5 species in the species list of outcomes)was 94%.The application results showed that it reduced the search range and identification time required for the fruit fly pests to a certain extent,partially solved the problem of their automatic identification in the quarantine or agriculture and forestry departments.