改进RetinaNet的农作物病虫害检测模型研究  

Research on Improving RetinaNet’s Crop Pest Detection Model

作　　者：邢伟寅 李礁[2] 钟乐海[2] 韩正勇 XING Weiyin;LI Jiao;ZHONG Lehai;HAN ZhengYong(Graduate School of JoséRizal University,Mandaluyong 1552,Manila,Philippines;School of Electronics and Information,Mianyang Polytechnic,Mianyang 621000,Sichuan,China)

机构地区：[1]菲律宾黎刹大学研究生院,马尼拉曼达卢永1552 [2]绵阳职业技术学院电子与信息学院,四川绵阳621000

出　　处：《四川农业大学学报》2023年第1期153-157,184,共6页Journal of Sichuan Agricultural University

基　　金：四川省科技计划重点研发项目(2022YFG0206);四川省级知识产权专项资金项目(2022-ZS-00156)。

摘　　要：【目的】为快速准确检测密集遮挡环境下农作物病虫害情况,满足大规模、高效率的识别需求。研究提出了一种改进RetinaNet的农作物病虫害检测模型。【方法】结合DenseNet改造RetinaNet的特征提取网络,强化特征重用,提高深度网络模型对农作物害虫的识别率,准确、快速地识别出病虫害的种类。其非极大抑制算法采用了DIoU策略,有效降低了在目标密集的情况下农作物病虫害的漏检率,提高了检测精度。【结果】改进后的算法模型具有较高的检测精度和良好的实时性,对作物密集遮挡情况有较好的适应性,mmAP达到了49.8%。【结论】将模型应用于复杂的农作物生长环境中,能有效提升病虫害智能检测能力。【Objective】 To quickly and accurately detect crop pests and diseases in dense shading environments to meet the needs of large-scale and high-efficiency identification. The study proposes an improved RetinaNet crop pest detection model. 【Method】 Combines with DenseNet, the RetinaNet feature extraction network was modified to enhance feature reuse, improve the recognition rate of the deep network model for crop pests, and identify the species of pests and diseases accurately and quickly. Its nongreat suppression algorithm adopts the DIoU strategy, which effectively reduces the leakage rate of crop pests in the case of dense targets and improves the detection accuracy. 【Result】 The experimental results show that the improved algorithm model has high detection accuracy and good real-time performance, and has good adaptability to the dense crop shading situation, and the mmAP reaches 49.8%.【Conclusion】 Applying the model to the complex crop growth environment can effectively improve the intelligent detection of pests and diseases.

关 键 词：病虫害识别 深度学习 目标检测 RetinaNet 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]