基于CNN的作物分类识别图像获取平台研究进展  

Research progress of image acquisition platform for crop classification and recognition based on CNN

作　　者：张倩[1] 王明 于峰[1] 陶震宇 张辉[1] 李刚 Zhang Qian;Wang Ming;Yu Feng;Tao Zhenyu;Zhang Hui;Li Gang(Institute of Data Science and Agricultural Economics,Beijing Academy of Agriculture and Forestry Sciences,Beijing,100097,China;Beijing Agricom Network Technology Co.,Ltd.,Beijing,100070,China)

机构地区：[1]北京市农林科学院数据科学与农业经济研究所,北京市100097 [2]北京智农天地网络技术有限公司,北京市100070

出　　处：《中国农机化学报》2024年第8期170-179,共10页Journal of Chinese Agricultural Mechanization

基　　金：北京市数字农业创新团队(BAIC10—2023);北京市农林科学院青年基金(QNJJ202213);北京市农林科学院改革与发展项目(GGFZSJS2024)。

摘　　要：基于机器视觉的作物精准分类识别是农业自动化、智能化作业的前提。在作物图像分类识别任务中,卷积神经网络(CNN)是当前应用最广泛的算法之一。作物表型特征及生长环境的复杂性,决定作物图像获取平台的多样性。通过分析2020—2022年国内外基于CNN的作物分类识别研究,图像获取平台可划分为通用平台和自建平台两大类:通用平台硬件产品成熟、部署方便,但要做好设备选型和环境搭建;自建平台分为固定式和移动式,能高效获取试验数据,但硬件集成较为复杂。详细对比分析各类平台的优缺点及适用范围。作物图像获取平台的未来趋势包括:高通量、高效率、自动化的通用图像获取装置,集成多种传感器的多模态数据采集与融合应用,自带运算处理的智能摄像头等,更精细化的图像获取平台将有效支撑作物表型的深入研究。Accurate crop classification and recognition based on machine vision is the premise of agricultural automation and intelligent operation.Convolution neural network(CNN)is one of the most widely used algorithms in crop image classification and recognition.The complexity of crop phenotypic characteristics and growth environment determines the diversity of crop image acquisition platforms.Through the analysis of crop classification and recognition research based on CNN at home and abroad from 2020 to 2022,the image acquisition platforms can be divided into two categories such as general platform and self⁃built platform,among which the general platform hardware products are mature and easy to deploy,but equipment selection and environment construction should be done well.The self⁃built platform is divided into fixed and mobile ones,which can obtain experimental data efficiently,but the hardware integration is more complicated.The advantages and disadvantages of various platforms and their applicable scope are compared and analyzed in detail.The future trends of crop image acquisition platforms will include high⁃throughput,high⁃efficiency,automated universal image acquisition devices,multi⁃modal data acquisition and fusion applications integrating a variety of sensors,intelligent cameras with built⁃in computing processing,etc.The more refined image acquisition platforms will effectively support in⁃depth research on crop phenotypes.

关 键 词：作物表型 机器学习 卷积神经网络 图像获取 作物分类识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP212[自动化与计算机技术—计算机科学与技术]