昆虫学报 ›› 2025, Vol. 68 ›› Issue (6): 816-829.doi: 10.16380/j.kcxb.2025.06.013

• 研究论文 • 上一篇    下一篇

一种新型移动计算场景下的高精度蚜虫识别与计数方法

宋一泓1, 王晨晞1, 张菁娟1, 包可晗1, 谭晶灵1, 张心阳1龚浩然1, 刘昱菲1, 张显2, 闫硕1,*   

  1. (1. 中国农业大学植物保护学院, 北京 100193; 2. 蒙自市蒙生石榴产销专业合作社, 蒙自 661100)
  • 出版日期:2025-06-20 发布日期:2025-07-31

A novel high-precision method for aphid recognition and counting in mobile computing scenarios

SONG Yi-Hong1, WANG Chen-Xi1, ZHANG Jing-Juan1, BAO Ke-Han1, TAN Jing-Ling1, ZHANG Xin-Yang1, GONG Hao-Ran1, LIU Yu-Fei1, ZHANG Xian2, YAN Shuo1,*   

  1.  (1. College of Plant Protection, China Agricultural University, Beijing 100193, China; 2. Mengzi City Mengsheng Pomegranate Production and Marketing Specialized Cooperative, Mengzi 661100, China)
  • Online:2025-06-20 Published:2025-07-31

摘要: 【目的】为减轻基层植保人员工作量,实现蚜虫准确和实时监控,建立了一种新型蚜虫识别方法。【方法】基于优化后的Transformer模型和稀疏注意力机制,开发了一种新的蚜虫识别和计数方法,特别适用于边缘计算环境。针对4种常见蚜虫——桃蚜Myzus persicae、棉蚜Aphis gossypii、豌豆蚜Acyrthosiphon pisum和禾谷缢管蚜Rhopalosiphum padi,利用相机HDV-56003与手机镜头构建数据采集系统;利用微信小程序展示蚜虫检测与计数结果;利用消融实验验证系统的可行性。【结果】本方法在准确率(accuracy)、平均精度均值(mean average precision, mAP)和每秒帧数(frames per second, FPS)等关键性能指标上均达到了优异的水平。具体而言,准确率达到了98%, mAP达到了95%,而FPS达到了52.3,这些指标均优于传统方法和其他基线模型。此外,本实验还开发了相应的移动应用程序,使农业从业者能够在田间环境中直接借助智能手机完成蚜虫的实时识别和计数,极大地提升了操作的便捷性和效率。【结论】基于Transformer的移动计算场景下的高精度蚜虫识别与计数方法实现了实时蚜虫识别和计数。本研究不仅为蚜虫精准识别提供了新型技术解决方案,也为其他农业害虫的智能化识别与监测提供了重要参考,有助于推动精准农业和智慧农业的发展。

关键词:  蚜虫, 蚜虫识别, 蚜虫计数, 农业害虫监测, 深度学习, 移动计算

Abstract: 【Aim】 To reduce the workload of grassroots plant protection personnel and realize the accurate and real-time monitoring of aphids, a new aphid recognition method has been established. 【Methods】 A new aphid recognition and counting method, which is particularly suitable for edge computing environments, was developed based on the optimized Transformer model and sparse attention mechanism. A data acquisition system was constructed using a camera HDV-56003 and a mobile phone lens for four common aphid species, Myzus persicae, Aphis gossypii, Acythosiphon pisum and Rhopalosiphum padi. WeChat mini program was used to display aphid detection and counting results, and ablation experiments were done to verify the feasibility of the system. 【Results】 The present method achieved excellent levels in key performance indicators such as accuracy, mean average precision (mAP), and frames per second (FPS). Specifically, the accuracy, mAP and FPS of the present method reached 98%, 95% and 52.3, respectively, all of them were superior to those of the traditional methods and other baseline models. In addition, in this experiment a corresponding mobile application program was also developed, allowing agricultural practitioners to directly complete the real-time recognition and counting of aphids in the field environment with the help of smartphones, and greatly improving the convenience and efficiency of operation. 【Conclusion】 The high-precision aphid recognition and counting method based on Transformer in mobile computing scenarios has achieved real-time aphid recognition and counting. This study not only provides new technological solutions for precise recognition of aphids, but also provides important references for the intelligent recognition and monitoring of other agricultural pests, which helps to promote the development of precision agriculture and smart agriculture.

Key words: Aphid, aphid recognition, aphid counting, agricultural pest monitoring, deep learning, mobile computing