基于深度学习的实蝇图像智能识别方法的建立

实蝇科昆虫很多种类是世界性的重要检疫性有害生物。实蝇的快速准确识别对于保护国门生物安全、促进我国农产品出口具有重要意义。本研究建立了一种基于深度学习的实蝇图像智能识别方法,针对实蝇科昆虫的翅图像特征,以桔小实蝇、南瓜实蝇、瓜实蝇、具条实蝇4种实蝇(每种250头共1000头)为例,通过标本制作与图像采集、图像预处理、创建数据集训练物体检测模型等模块不断调整优化实现对实蝇图像的自动识别。结果表明在图像色阶参数90/1/220,通过百度Easy DL平台Paddle Paddle深度学习框架结合Auto Model Search训练图像分类模型,选择超高精度算法、高级训练配置epoch并使用数据增强策略进行模型训练时识别准确率达95%以上。本方法具有操作简单、准确率高、可扩展性强等特性,通过智能手机拍摄待测样本输入系统即可进行准确识别,可应用于果蔬园实蝇监测、出入境口岸实蝇检疫以及昆虫科普教育等场景,并可为其他昆虫自动识别研究提供有益借鉴。

Establishment of fruit fly intelligent image identification based on deep learning

Many species of Tephritidae are important quarantine pests in the world.The rapid and accurate identification of fruit flies is of great significance for the protection of national biosafety and the promotion of agricultural exports.In this study,an intelligent identification method of fruit fly image based on deep learning was established.According to the wing image characteristics of fruit flies,taking Bactrocera dorsalis,B.cucurbitae,B.tau and B.scutellate as examples(250 samples for each species,1000 in total),the automatic identification method of fruit fly image was developed by adjusting and optimizing the modules of specimen making and image acquisition,image preprocessing,creating data set and training object detection model.The results showed that the identification accuracy was more than 95%when the image color order parameter was 90/1/220,the super-high-precision algorithm was selected,the epoch was configured in advanced training and the data enhancement strategy was used for model training through Paddle Paddle deep learning framework and Auto Model Search of Baidu’s Easy DL platform.This method has the characteristics of simple operation,high accuracy and strong expansibility.We only took the sample photos by smart phone,and then input them into the system for accurate recognition of fruit fly species.It can be used for monitoring of fruit flies in fruit and vegetable orchards,quarantine at entry-exit ports and education of insect science popularization.Moreover,it can also provide useful reference for other insect automatic identification research.