基于YOLOv8-FECA的密集场景下小麦赤霉病孢子目标检测  

Detection of wheat scab spores in dense scene based on YOLOv8-FECA

作　　者：张东彦 高玥[2] 程涛 胡根生 杨雪[3] 乔红波 郭伟 谷春艳[3] ZHANG Dongyan;GAO Yue;CHENG Tao;HU Gensheng;YANG Xue;QIAO Hongbo;GUO Wei;GU Chunyan(College of Mechanical and Electronic Engineering,Northwest A&F University,Yangling 712100,China;National Engineering Research Center for Agro-Ecological Big Data Analysis&Application,Anhui University,Hefei 230601,China;Institute of Plant Protection and Agro-products Safety,Anhui Academy of Agricultural Sciences,Hefei 236065,China;College of Information and Management Science,Henan Agricultural University,Zhengzhou 450002,China)

机构地区：[1]西北农林科技大学机械与电子工程学院,杨凌712100 [2]安徽大学,农业生态大数据分析与应用技术国家地方联合工程研究中心,合肥230601 [3]安徽省农业科学院植物保护与农产品质量安全研究所,合肥236065 [4]河南农业大学信息与管理科学学院,郑州450002

出　　处：《农业工程学报》2024年第21期127-136,共10页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然基金面上项目(42271364);河南省重点研发专项(241111110800);安徽省农业科技成果转化应用专项(2024ZH004)。

摘　　要：针对小麦赤霉病孢子图像中密集分布小目标的漏检错检问题,该研究设计了一种针对该场景下小目标检测模型YOLOv8-FECA。以YOLOv8为基准模型,先添加新的小目标检测层,以此增强网络对更小目标语义信息的捕捉以及提高其特征描述的准确性;其次,构建融合特征的焦点注意力机制(focal efficient channel attention module,FECA)以实现对小目标和密集分布目标的关注;再次,引入Wise-IoU Loss与DFL Loss结合作为边界框的回归损失,提高模型的收敛能力以及对边界框预测的准确性;最后,在不同密集场景和不同光线环境下验证了模型的鲁棒性。结果表明,优化后模型YOLOv8-FECA相比YOLOv8在孢子数据集中的平均检测精度均值m AP@0.5提高了4.3个百分点,达到96.8%,且改进模型的检测效率达到实时检测要求。该研究设计的YOLOv8-FECA网络模型在保持高性能的同时能够有效提升密集场景小目标检测的精度,为其他作物病害孢子的检测提供了思路。Wheat scab is one of the most serious diseases that threaten global wheat production.The primary causal agents can be from the fungi Fusarium graminearum and Fusarium asiaticum.The Fusarium spores can be disseminated in the transmission and infection of wheat scab.Consequently,the early and precise identification of pathogen spore concentrations is of paramount importance to prompt the detection of wheat scab,particularly for the high wheat yield and quality,as well as food safety.Conventional techniques,such as microscopic observation and medium isolation,have posed a great challenge to the rapid response and the transmission of large-scale crop disease,due to time-consuming and sophisticated technical expertise.Furthermore,there are missed and erroneous detections of densely distributed small targets in scab spore images.In this study,a small target detection model,YOLOv8-FECA was developed to overcome this challenge.Firstly,a small target detection layer with a feature map size of 160x160 was added to the original YOLOv8 baseline model.The network was then improved to capture the semantic information of small targets,in order to enhance the accuracy of feature description.Secondly,a focus attention mechanism,FECA,was designed to combine the CBAM and ECA modules.This module was then added at the neck end of the network to reduce information loss during downsampling.Ultimately,Wise-IoU Loss and DFL Loss were integrated as the regression loss for the bounding box,in order to enhance the convergence and the precision of bounding box estimation.The experimental results demonstrate that the small target detection layer was added to introduce the focus attention mechanism,compared with the original YOLOv8 baseline model.Wise-IoU Loss and DFL were also introduced after Loss improvement.The average detection accuracy of the new YOLOv8-FECA on spore data set mAP@0.5 and mAP@0.5:0.95 increased by 4.3 percentage points and 6.3 percentage points,respectively,compared with the benchmark model Yolov8n.In terms of performance

关 键 词：小麦赤霉病 孢子 密集场景 小目标检测 YOLOv8 显微图像 

分 类 号：S126[农业科学—农业基础科学]