基于改进 YOLOX 模型的柑橘木虱检测方法

王海漫 1; 2; 俞　婷 1; 肖明明 3; 杨嘉诚 1; 2; 陈富荣 2; 易干军 4; 林德球 5; 罗　敏 5

文章摘要

王海漫 1, 2，俞　婷 1 ，肖明明 3 ，杨嘉诚 1, 2，陈富荣 2 ，易干军 4 ，林德球 5 ，罗　敏 5.基于改进 YOLOX 模型的柑橘木虱检测方法[J].广东农业科学,2022,49(11):43-49

查看全文 HTML 基于改进 YOLOX 模型的柑橘木虱检测方法

Detection of Citrus Psyllid Based onImproved YOLOX Model

DOI：10.16768/j.issn.1004-874X.2022.11.005

中文关键词: 柑橘改进 YOLOX 模型木虱防控人工智能目标检测

英文关键词: citrus improved YOLOX model prevention and control of psyllid artificial intelligence object detection

基金项目:广东省重点领域研发计划项目（2020B0202090005）；廉江智库企业项目（廉江红橙黄龙病与木虱智慧监控及生态防控技术应用示范）

作者

单位

王海漫 1, 2，俞　婷 1 ，肖明明 3 ，杨嘉诚 1, 2，陈富荣 2 ，易干军 4 ，林德球 5 ，罗　敏 5

1. 广东省农业科学院农业生物基因研究中心 / 广东省农作物种质资源保存与利用重点实验室，广东广州　510640；2. 仲恺农业工程学院信息科学与技术学院，广东广州　510225； 3. 广州航海学院信息与通信工程学院，广东广州　510725； 4. 广东省农业科学院，广东广州 510640； 5. 廉江市经济社会发展研究会，广东廉江　524400

摘要点击次数: 812

全文下载次数: 630

中文摘要:

【目的】黄龙病被称为柑橘的“癌症”，是一种毁灭性病害，而木虱是黄龙病传播的主要媒介，对木虱的监测和精准消杀是防控黄龙病及抑制其传播的一种有效途径。【方法】传统方式消灭木虱主要是靠人工喷洒药物，人力成本高但防控效果并不理想。采用基于改进 YOLOX 的木虱边缘检测方法，在主干网络加入卷积注意力模块 CBAM（Convolutional block attention module），在通道和空间两个维度对重要特征进行进一步提取；将目标损失中的交叉熵损失改为使用 Focal Loss，进一步降低漏检率。【结果】本研究设计的算法契合木虱检测平台，木虱数据集拍摄于广东省湛江市廉江红橙园，深度适应农业农村实际发展需要，基于 YOLOX 模型对骨干网络和损失函数做出改进实现了更加优秀的柑橘木虱检测方法，在柑橘木虱数据集上获得 85.66% 的 AP 值，比原始模型提升 2.70 个百分点，检测精度比 YOLOv3、YOLOv4-Tiny、YOLOv5-s 模型分别高 8.61、4.23、3.62 个百分点，识别准确率大幅提升。【结论】改进的 YOLOX 模型可以更好地识别柑橘木虱，准确率得到提升，为后续实时检测平台打下了基础。

英文摘要:

【Objective】Yellow-shoot disease, known as the cancer of citrus, is a devastating disease, and psyllid is the main vector of yellow-shoot disease transmission, therefore, monitoring and precise disinfection and sterilization of psyllid is an effective way to prevent and control yellow-shoot disease and inhibit its transmission.【Method】The traditional way to eliminate the psyllid was mainly to spray drugs manually, and the control effect was not ideal due to high labor costs. In the study, we used an improved YOLOX based edge detection method for psyllid, added Convolutional Block Attention Module (CBAM) to the backbone network, and further extracted important features in the channel and space dimensions. The cross entropy loss in the target loss was changed to Focal Loss to further reduce the missed detection rate.【Result】The results showed that the algorithm described in the study fitted in with the detection platform of psyllid. The data set of psyllid was taken in Lianjiang Orange Garden, Zhanjiang City, Guangdong Province. It is deeply adapted to the actual needs of agricultural and rural development. Based on YOLOX model, the backbone network and loss function were improved to achieve a more excellent detection method of citrus psyllid. 85.66% of the AP value was obtained on the data set of citrus psyllid, which was 2.70 percentage points higher than that of the original model, and the detection accuracy was 8.61, 4.32 and 3.62 percentage points higher than that of YOLOv3, YOLOv4-Tiny and YOLOv5-s respectively, which has been greatly improved.【Conclusion】The improved YOLOX model can better identify citrus psyllid, and the accuracy rate has been improved, laying a foundation for the subsequent real-time detection platform.

查看/发表评论下载PDF阅读器