基于GEB-YOLO v8n的烟草蚜虫识别与计数方法

Tobacco Aphid Identification and Counting Method Based on GEB-YOLO v8n

摘要:

蚜虫数量是直观表征烟草虫害侵染程度的表型指标之一，针对环境光照动态变化、图像模糊等田间图像采集问题，提出了一种GEB-YOLO v8n轻量化的烟草蚜虫检测算法。首先，主干网络引入GSConv与有效通道注意力机制(Efficient channel attention, ECA)，增强了协同输出烟蚜丰富的图像特征信息和目标导向能力；其次，颈部网络引入加权双向特征金字塔网络(Bidirectional feature pyramid network, BiFPN)，增强了模型检测烟蚜特征图的语义表达能力和空间信息质量；最后，引入WIoU作为边界框回归损失函数，通过动态聚焦复杂样本，使模型更好地泛化到新的、挑战性强的烟蚜检测场景。经模型结构重参数化与超参数优化后，形成面向田间烟蚜检测的网络架构。结果表明，改进后的模型平均精度均值(Mean average precision，mAP)和F1值分别达到了91.8%和90.4%，参数量(Parameters)降低42.8%，模型内存占用量(Model memory footprint)和浮点运算次数（Floating point operations, FLOPs)分别降低至3.5MB和4.1×109，平均推理时间缩短至3.6ms。基于GEB-YOLO v8n模型开发了一款面向小区田间烟草蚜虫识别与计数系统，系统具备在线式图像检测与视频检测双功能，能够在界面上直观展示烟蚜数量检测结果，满足小区田间烟蚜实时检测以及移动端部署的需求。本研究轻量化GEB-YOLO v8n模型为田间环境中烟草植物病虫害识别与表型分析提供了方法参考。

Abstract:

A lightweight tobacco aphid detection algorithm called GEB-YOLO v8n was proposed to address the problems in field image acquisition, such as dynamic changes in ambient light and image blurring. Firstly, GSConv and the efficient channel attention (ECA) mechanism were innovatively introduced into the backbone network, and the rich image feature information and target-oriented ability of tobacco aphids were jointly output. Secondly, the bidirectional feature pyramid network (BiFPN) was introduced into the neck network, and the semantic expression ability and spatial information quality of the model for detecting tobacco aphid feature maps were enhanced. Finally, WIoU was introduced as the bounding-box regression loss function, and the model was enabled to better generalize to new and challenging tobacco aphid detection scenarios by dynamically focusing on complex samples. After the model structure re-parameterization and hyperparameter optimization, a network architecture for field tobacco aphid detection was formed. The results showed that the mean average precision (mAP) and F1 value of the improved model reached 91.8% and 90.4%, respectively, the number of parameters was reduced by 42.8%, the model memory footprint and floating point operations (FLOPs) were reduced to 3.5MB and 4.1×109, respectively, and the average inference time reached 3.6ms. A system for tobacco aphid recognition and counting in small-scale fields was developed based on the GEB-YOLO v8n model. The system had the dual functions of online image detection and video detection, and can intuitively display the detection results of the number of tobacco aphids on the interface, meeting the requirements of real-time detection of tobacco aphids in small-scale fields and mobile-end deployment. The improved lightweight GEB-YOLO v8n model can provide a method reference for the identification and phenotypic analysis of tobacco plant diseases and pests in the field environment.