一种基于GhostNet的绿色类圆果实识别方法  

A green round-like fruits identification method based on GhostNet

作　　者：李恒 南新元[1] 高丙朋[1] 马志钢 LI Heng;NAN Xin-yuan;GAO Bing-peng;MA Zhi-gang(School of Electrical Engineering,Xinjiang University/Siemens Laboratories,Urumqi 830017,China)

机构地区：[1]新疆大学电气工程学院/西门子实验室,新疆乌鲁木齐830017

出　　处：《江苏农业学报》2023年第3期724-731,共8页Jiangsu Journal of Agricultural Sciences

基　　金：国家自然科学基金项目(61863033)。

摘　　要：为实现果园实际环境中绿色类圆果实的识别,研究了基于单阶段目标检测网络的绿色类圆果实识别方法。本研究对比4种不同轻量化卷积网络模型,以GhostNet作为本研究网络的主干特征提取网络,将提取到的特征信息利用复杂双向多尺度融合网络进行融合,最后以改进后的YOLO_Head作为预测头,建立适合本研究的目标检测网络。结果表明,在果园背景下本研究构建的目标检测网络对绿色类圆果实的均值平均精度达到96.8%,每张图片检测所用的时间为37 ms,网络内存占用大小为11.8 M,实现了对绿色类圆果实的快速、准确识别,能够为早期果树的产量预估、病虫害识别提供技术支撑。In order to realize the recognition of green round-like fruits in the actual environment of orchards,the recognition method of green round-like fruits based on one-stage object detection network was studied.In this study,four different lightweight convolutional network models were compared.GhostNet was used as the backbone feature extraction network of this research network.The extracted feature information was fused by bidirectional feature pyramid network(BiFPN).Finally,the improved YOLO_Head was used as the prediction head to establish a target detection network suitable for this study.The experimental results showed that the final detection accuracy of the green round-like fruits in the object detection network constructed in the context of orchard reached 96.8%,the detection speed of a single image reached 37 ms,and the memory occupancy size of the network was 11.8 M,which realized the rapid and accurate identification of green round-like fruits,and could provide technical support for the yield estimation and disease and pest identification of early fruit trees.

关 键 词：目标检测 轻量化卷积网络 特征融合 绿色类圆果实 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]