基于深度学习的棉花原位根系表型的识别  

Identification of cotton in-situ root phenotypes based on deep learning

作　　者：李安昌 于秋实 徐文君[2] 祝令晓 滕桂法[4] LI Anchang;YU Qiushi;XU Wenjun;ZHU Lingxiao;TENG Guifa(College of Mechanical&Electrical Engineering,Hebei Agricultural University,Baoding 071001,China;College of Science,Hebei Agricultural University,Baoding 071001,China;College of Agronomy,Hebei Agricultural University,Baoding 071001,China;College of Information Science and Technology,Hebei Agricultural University,Baoding 071001,China)

机构地区：[1]河北农业大学机电工程学院,河北保定071001 [2]河北农业大学理学院,河北保定071001 [3]河北农业大学农学院,河北保定071001 [4]河北农业大学信息科学与技术学院,河北保定071001

出　　处：《河北农业大学学报》2024年第2期17-21,共5页Journal of Hebei Agricultural University

基　　金：河北省教育厅青年拔尖人才项目(BJ2021058)。

摘　　要：原位根系研究是探测根系表型及变化动态的重要方法,被广泛应用。然而,传统根系图像分割手段效率低、精度差,是制约根系研究的关键障碍。为实现原位根系图像分割的高效和高精度,本文基于语义分割U-Net网络设计与优化,在跳跃链接中加入SE模块,替换优化器为Lion,实现原位根系表型的精准识别。进一步,采用1D-CNN网络,对原位根系表型信息进行特征挖掘。验证结果显示,相较于原始模型,改进后的U-Net在精度上提高了1.57%,交并比提高了3.41%;1D-CNN对表型参数鉴定的精度为90.9%。本研究基于深度学习方法,实现了原位根系的高效和精准识别与分割,为棉花原位根系研究提供了重要支撑。In situ root system research is an important method for exploring root system morphology and dynamic changes,and it has been widely applied.However,traditional methods for image segmentation of root systems suffer from low efficiency and poor accuracy,which are key obstacles to in situ root system research.To achieve efficient and accurate segmentation of in situ root system images,this paper designed and improved a U-Net network based on semantic segmentation.SE modules were incorporated in the skip-connection,and the optimizer was replaced with Lion,enabling precise identification of in situ root system phenotype.Furthermore,a 1D-CNN network was employed to extract features of phenotypic information from the in situ root system.The validation results showed that the improved U-Net achieved a 1.57%increase in accuracy and a 3.41%improvement in intersection over union(IoU)compared with that of the original model.The identification accuracy of phenotype parameter using 1D-CNN was 90.9%.This study realized efficient and accurate identification and segmentation of in situ root systems through deep learning method,providing important support for in situ root system research in cotton.

关 键 词：棉花原位根系 表型识别 改进U-Net 1D-CNN 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]