融合注意力机制的荔枝轻量化检测方法研究  

A Light Weight Detection Method of Litchi by Fusion of Attention Mechanism

作　　者：王聪 文晟[1,2] 兰玉彬 严倩 姜永华 张建桃 罗菊川 Wang Cong;Wen Sheng;Lan Yubin;Yan Qian;Jiang Yonghua;Zhang Jiantao;Luo Juchuan(College of Engineering,South China Agricultural University,Guangzhou 510642,China;National Joint Research Center of Precision Agriculture Aviation Application Technology,Guangzhou 510642,China;College of Electronic Engineering/College of Artificial Intelligence,South China Agricultural University,Guangzhou 510642,China;Fruit Research Institute,Guangdong Academy of Agricultural Sciences,Guangzhou 510645,China;College of Mathematics and Information,South China Agricultural University,Guangzhou 510642,China)

机构地区：[1]华南农业大学工程学院,广州510642 [2]国家精准农业航空施药技术国际联合研究中心,广州510642 [3]华南农业大学电子工程学院/人工智能学院,广州510642 [4]广东省农业科学院果树研究所,广州510645 [5]华南农业大学数学与信息学院,广州510642

出　　处：《农机化研究》2025年第3期10-15,共6页Journal of Agricultural Mechanization Research

基　　金：国家自然科学基金面上项目(32271985);广东省自然科学基金面上项目(2022A1515011008)。

摘　　要：针对荔枝果实个体小、生长密集和遮挡严重等特点,为了快速准确地实现荔枝的检测和计数,提出了一种融合注意力机制和多尺度特征图的网络模型。为了提高模型对遮挡和阴影环境下果实的识别准确率,将Coordinate Attention(CA)注意力机制嵌入至YOLOv4-Tiny模型。为了提高模型对小目标果实的检测精度,在特征金字塔Feature Pyramid Networks(FPN)结构中生成了两个更大尺度的特征图。试验结果表明:融合注意力机制的荔枝轻量化检测模型的准确率(Precision)、召回率(Recall)和平均精度(mAP)分别为92.92%、76.09%和88.51%。与YOLOv4-Tiny和YOLOv3模型相比,所构建的融合注意力机制的荔枝轻量化检测模型的平均检测精度分别高出8.84个百分点和3.87个百分点。该模型能够快速、精准地检测出果园环境中的荔枝,适用于果园中荔枝的识别和计数。In view of the characteristic of small size,dense growth and severe occlusion of litchi,in order to rapidly and accurately detect and count litchi fruits,a network model combining attention mechanism and multi-scale feature maps was proposed in this study.For the purpose of improving the recognition accuracy of fruit in occluding and shaded environments,Coordinate Attention(CA)mechanism was embedded in YOLOv4-Tiny model.In order to improve the detection accuracy of the model for small target fruit,two larger scale feature maps were generated in the structure of Feature Pyramid Networks(FPN).The results showed that the Precision,Recall and mAP of the Litchi lightweight detection model combined with attention mechanism were 92.92%,76.09%and 88.51%,respectively.Compared with YOLOv4-Tiny and YOLOv3 models,the average detection accuracy of litchi lightweight detection model constructed in this paper with integrated attention mechanism was 8.84 percentage points and 3.87 percentage points higher,respectively.The model can detect litchi quickly and accurately in orchard environment,and suitable for identification and counting of litchi in orchard.

关 键 词：荔枝 注意力机制 特征金字塔 轻量化 检测方法 

分 类 号：S126[农业科学—农业基础科学] TP391.4[自动化与计算机技术—计算机应用技术]