基于改进YOLOv8的自然环境下板栗栗蓬检测方法  

Detection method of Chinese chestnut in natural environment based on improved YOLOv8

作　　者：李志臣 罗卫平[1] 凌秀军[1] 李鸿秋[1] Li Zhichen;Luo Weiping;Ling Xiujun;Li Hongqiu(School of Mechanical&Electrical Engineering,Jinling Institute of Technology,Nanjing,211169,China)

机构地区：[1]金陵科技学院机电工程学院,南京市211169

出　　处：《中国农机化学报》2024年第12期251-258,共8页Journal of Chinese Agricultural Mechanization

基　　金：国家自然科学基金面上项目(51775270)。

摘　　要：针对人工敲打式收获板栗带来高成本和安全风险问题,研究无人机板栗采摘方法至关重要。为达到在自然光照条件下迅速且精确地识别板栗栗蓬目标,提出一种基于YOLOv8的改进卷积网络模型板栗栗蓬检测方法。对YOLOv8主干网络的C2f模块添加CBAM注意力机制,增强卷积网络模型对板栗栗蓬特征提取能力。在YOLOv8的头部增加一个微小栗蓬目标检测头,与YOLOv8原有的3个检测头共同组成检测模块,使网络模型更好地捕捉小板栗栗蓬目标特征。经自建数据集上的训练和验证试验,改进后卷积网络YOLOv8-Vcj板栗栗蓬检测精确率比YOLOv8高1.3%,mAP@0.5和mAP@0.5∶0.95值比YOLOv8分别提高4.6%和3.4%。改进卷积网络板栗栗蓬检测误差主要来自光照条件和图像中板栗栗蓬目标的密集程度。研究结果表明:融合CBAM注意力机制和增加微小目标检测头的改进卷积神经网络YOLOv8-Vcj能够有效实现树上板栗栗蓬的检测。In view of the high cost and safety risk caused by artificial knock harvest chestnut,it is very important to study the unmanned aerial vehicle chestnut harvest methods.In order to rapidly and precisely identify chestnut targets under natural light conditions,a modified convolutional network model detection method based on YOLOv8 was proposed.The CBAM attention mechanism was added to the C2f module of the YOLOv8 backbone network to enhance the convolutional network model ability of extracting chestnut features.A small chestnut target detection head was added to the head of YOLOv8 which formed the detection module together with the original three detection heads of YOLOv8.This method(YOLOv8-Vcj)enabled the network model to better capture the target features of small chestnut.Through training and validation experiments on the self-built data set,the detection accuracy of YOLOv8-Vcj was 1.3%higher than YOLOv8 and the mAP@0.5 and mAP@0.5∶0.95 values were 4.6%and 3.4%higher than YOLOv8,respectively.The chestnut detection error of the improved convolution network mainly comes from the light conditions and the density of chestnut targets in the images.The research results show that the improved convolutional neural network YOLOv8-Vcj of combining the CBAM attention mechanism and a small target detection head can effectively detect chestnuts on the tree.

关 键 词：板栗栗蓬 YOLOv8 目标检测 CBAM 检测头 

分 类 号：TS736.4[轻工技术与工程—制浆造纸工程]