姜红花,王鹏飞,张 昭,毛文华,赵 博,齐 鹏.基于卷积网络和哈希码的玉米田间杂草快速识别方法[J].农业机械学报,2018,49(11):30-38.
JIANG Honghua,WANG Pengfei,ZHANG Zhao,MAO Wenhua,ZHAO Bo,QI Peng.Fast Identification of Field Weeds Based on Deep Convolutional Network and Binary Hash Code[J].Transactions of the Chinese Society for Agricultural Machinery,2018,49(11):30-38.
摘要点击次数: 2022
全文下载次数: 938
基于卷积网络和哈希码的玉米田间杂草快速识别方法   [下载全文]
Fast Identification of Field Weeds Based on Deep Convolutional Network and Binary Hash Code   [Download Pdf][in English]
投稿时间:2018-05-17  
DOI:10.6041/j.issn.1000-1298.2018.11.004
中文关键词:  杂草识别  卷积神经网络  哈希码  深度学习  特征压缩
基金项目:山东省重点研发计划项目(2015GNC112004)、国家重点研发计划项目(2017YFD0700500)、山东省自然科学基金项目(ZR2018MC017)和山东农业大学智能化农业装备研发项目(24132)
作者单位
姜红花 山东农业大学 
王鹏飞 山东农业大学 
张 昭 宝鸡文理学院 
毛文华 中国农业机械化科学研究院 
赵 博 中国农业机械化科学研究院 
齐 鹏 山东永佳动力股份有限公司 
中文摘要:为提高作物与杂草识别的准确性,结合深度卷积网络强大的特征提取能力和哈希码便于存储和快速检索的特点,提出了基于深度卷积网络和二进制哈希码的田间杂草快速识别方法。结合预训练的多层卷积网络,增加二进制哈希层构建杂草识别模型,并利用所采集的杂草数据集对模型进行fine-tuning。所提出的二进制哈希层可有效地将高维杂草特征进行压缩,以便于实际田间杂草特征的存储和后续计算。在进行杂草识别时,利用训练好的模型提取输入图像的全连接层特征码和哈希特征码,与数据库中的全连接层特征码和哈希特征码进行对比,分别计算其汉明距离与欧氏距离,找出与其最相似的K幅图像,统计这K幅图像的标签,将其归入频率最高的一类,以达到分类识别的目的。通过对比不同卷积层数和不同二进制哈希码长度对杂草识别的影响,最终确定了包含4层卷积网络和128位哈希码长度的杂草识别模型。试验结果表明,本研究方法田间杂草识别准确率可达98.6%,并且损失函数稳定性相较于普通模型有所提高;同时,在其他杂草数据集上也有良好的表现,准确率达到95.8%,说明该方法具有通用性。实地测试表明,利用本文提出的模型进行杂草识别,对靶喷雾杂草施药率可达92.7%,能够有效减少农药浪费,适用于精准喷雾。
JIANG Honghua  WANG Pengfei  ZHANG Zhao  MAO Wenhua  ZHAO Bo  QI Peng
Shandong Agricultural University,Shandong Agricultural University,Baoji University of Arts and Sciences,Chinese Academy of Agricultural Mechanization Sciences,Chinese Academy of Agricultural Mechanization Sciences and Shandong Yongjia Power Co., Ltd.
Key Words:weed identification  convolution neural network  Hash code  deep learning  feature compression
Abstract:Corn is one of the main grain crops in China and its production accounts for more than 20% of the World’s corn production. Weed is one of the most important factors influencing maize yield. Effective recognition method of cron and weed can improve corn quality and production accounts. At present, pesticide spraying is the main way of removing weed in China. Excessive spraying of pesticides brings problems such as environmental pollution and food safety, and therefore precise spraying is the key of weeding to reduce the amount of pesticides and increase the utilization of pesticides. Precise application of pesticides is based on accurate identification of weeds, researchers at home and abroad have done a lot of research. Most existing weed identification methods rely on manually selected weed features, such as shape, texture, etc., which takes longer time to identify the image, and the accuracy of identification still needs further improvement. The deep learning method was used to achieve automatic extraction of weed image features without relying on artificial feature screening, and combined the binary Hash code to compress high-dimensional weed feature data to achieve rapid weed identification and provide information support for subsequent field drug spraying. In order to improve accuracy of crop and weed identification, combining with the strong feature extraction capabilities of the deep convolutional network and the ease of storage and fast retrieval of the Hash code, a fast field weed identification method was proposed based on the deep convolutional network and binary Hash code. A pre trained multilayer convolutional neural network was used to construct a weed identification model with a binary Hash layer, and the model was fine-tuned with the collected weed data set. The binary Hash layer proposed could effectively compress the high dimensional weed image features to facilitate the storage and subsequent calculation of the high dimensional weed image features. During tests of weed identification, the trained model was used to extract the full connection layer feature codes and binary Hash codes of the input image, and then compared with the full connection layer feature codes and binary Hash codes stored in the database to calculate the Hamming distance and the Euclidean distance. After that, the most similar K images could be found out according to last step’s results. Finally, the labels’ frequency of the K images was counted and the original image was classified into the highest frequency category of label to achieve the purpose of weed identification. The effects of different layers of convolutional networks and different length binary Hash code on weed identification were compared, and finally the weed identification model was determined, which included four layers convolutional neural network and 128 bit binary Hash code. The experimental results showed that the method proposed could achieve 98.6% accuracy in field weed identification, and the loss function stability was improved compared with the ordinary model. At the same time, it also performed well on other weeds datasets with an accuracy of 95.8%, which meant that the proposed method was universal. The research results could provide reference for precision weeding. The experiment carried out in corn field showed that the method could achieve 92.7% accuracy, and it could effectively reduce pesticide waste which was suitable for precision spray.

Transactions of the Chinese Society for Agriculture Machinery (CSAM), in charged of China Association for Science and Technology (CAST), sponsored by CSAM and Chinese Academy of Agricultural Mechanization Science(CAAMS), started publication in 1957. It is the earliest interdisciplinary journal in Chinese which combines agricultural and engineering. It always closely grasps the development direction of agriculture engineering disciplines and the published papers represent the highest academic level of agriculture engineering in China. Currently, nearly 8,000 papers have been already published. There are around 3,000 papers contributed to the journal each year, but only around 600 of them will be accepted. Transactions of CSAM focuses on a wide range of agricultural machinery, irrigation, electronics, robotics, agro-products engineering, biological energy, agricultural structures and environment and more. Subjects in Transactions of the CSAM have been embodied by many internationally well-known index systems, such as: EI Compendex, CA, CSA, etc.

   下载PDF阅读器