引用本文:张文利,陈开臻,刘鈺昕,段玉林,郭 威,史 云※.基于边缘设备的轻量化小目标果实检测模型[J].中国农业信息,2021,33(1):28-36
【打印本页】   【HTML】   【下载PDF全文】   查看/发表评论  【EndNote】   【RefMan】   【BibTex】
←前一篇|后一篇→ 过刊浏览    高级检索
本文已被:浏览 26次   下载 5 本文二维码信息
码上扫一扫!
分享到: 微信 更多
基于边缘设备的轻量化小目标果实检测模型
张文利1,陈开臻1,刘鈺昕1,段玉林2,郭 威3,史 云※2
1.北京工业大学信息学部,北京100124;2.中国农业科学院农业资源与农业区划研究所,北京100081;3.日本东京大学田间表型实验室,东京188-0002
摘要:
【目的】随着计算机视觉和智慧农业的快速发展,果实检测技术已成为研究热点。然 而在果园实际应用场景中,存在模型计算量大、目标果实尺度小的问题,导致模型难以在边 缘设备上实时运行且小目标果实检测精度低,因此文章通过改进Yolov3 模型,设计并实现 一种轻量化小目标果实检测模型RegNet-Yolov3,能够在边缘设备上实时运行并实现高精度 果实检测。【方法】该模型通过构建轻量化特征提取网络,有效降低模型参数计算量,满足 在边缘设备上实时运行要求;并针对柑橘果实小尺度特点,通过添加浅层网络检测分支优 化模型小目标检测性能,提升检测精度。【结果】将模型部署在边缘设备Jetson TX2 nano 上 进行测试,模型mAP 值和网络推理速度分别为96.0% 和122 ms,均优于原先Yolov3 网络测 试结果。【结论】实验结果表明,该研究模型能够实现在保持较高检测精度下,在边缘设备 Jetson TX2 nano 上实时运行,满足果园作业平台果实检测工作。
关键词:  果实检测;Yolov3;小目标;轻量化;Jetson TX2 nano
DOI:10.12105/j.issn.1672-0423.20210103
分类号:
基金项目:中国农业科学院国际农业科学计划(CAAS-ZDRW202107)
Lightweight small target fruit detection model based on edge devices
Zhang Wenli1, Chen Kaizhen1, Liu Yuxin1, Duan Yulin2, Guo Wei3, Shi Yun※2
1.Department of Informatics,Beijing University of Technology,Beijing 100124,China;2.Institute of Agricultural Resources and Regional Planning,Chinese Academy of Agricultural Sciences,Beijing 100081,China;3.International Field Phenomics Research Laboratory,The University of Tokyo,Tokyo 188-0002,Japan
Abstract:
[Purpose]With the rapid development of modern computer technology and intelligent agriculture,computer vision technology has been widely used in agriculture,among which, fruit detection technology is one of the main research hotspots at present,and is the technical basis for intelligent orchard work such as fruit yield estimation,fruit positioning,and automatic fruit picking.In recent years,with the success of deep learning technology applied in other fields and the improvement of its ease of use,deep learning-based object detection technology has gradually replaced traditional object detection technology and is widely used in modern intelligent orchard fruit detection work.However,in practical orchard scenarios,many existing detection model parameters are computationally large,resulting in slow model inference,which makes it difficult to use on work platforms with limited hardware resources.And in the acquired images, the target fruit scale is usually smaller,which makes detection more difficult and leads to lower model detection accuracy.In order to be able to implement a running detection model on a working platform with limited hardware resources and to maintain high fruit detection accuracy,this paper designs and implements a lightweight small target fruit detection model,RegNet-Yolov3,by improving the Yolov3 model,which can run and detect target fruits in real time on edge devices. [Method]Based on the X_Block lightweight network module in RegNet network,the method redesigns and builds a lightweight feature extraction network,effectively reducing the amount of parameter calculations in the model,increasing the model inference speed,and meeting the requirements of real-time operation on the working platform with limited hardware resources.Since the detection of small-scale targets is mainly performed in the shallow layer of the network,the method optimizes the model’s detection performance for small-scale target fruits by removing the deepest network detection branch in the original Yolov3 and adding a shallow network detection branch.And the FPN network structure in the original Yolov3 model is used to effectively fuse deep and shallow feature information and improve fruit detection accuracy.[Result]Deploying the model on the edge device Jetson TX2 nano for testing,the mAP(Mean Average Precision) and network inference speed are 96.0% and 122 ms,respectively,which are better than the original Yolov3 network test results.[Conclusion]The experimental results show that the model is able to run on the edge device Jetson TX2 nano in real time while maintaining a high detection accuracy,which is suitable for fruit detection on orchard work platforms.
Key words:  fruit detection;Yolov3;small target;lightweight;Jetson TX2 nano