面向智慧农业的智能图像标注系统  

Intelligent image annotation system for smart agriculture

作　　者：张哲 李佩霏 吕宪勇 杨凯 ZHANG Zhe;LI Peifei;LÜXianyong;YANG Kai(Weichai Power Co.,Ltd.,Weifang 261000,Shandong,China)

机构地区：[1]潍柴动力股份有限公司,山东潍坊261000

出　　处：《农业装备与车辆工程》2024年第7期145-152,共8页Agricultural Equipment & Vehicle Engineering

基　　金：十四五国家重点研发计划项目(SQ2021YFB2500012)。

摘　　要：提出了一个面向智慧农业的智能图像标注系统,旨在解决智慧农业场景图像标注耗时问题。首先借鉴主动学习的思路,设计了全新的查询策略进行图片筛选,将待标注的图像分类为难标注图像和易标注图像,难标注图像用于优化自动标注模型,有利于提高自动标注的精度,为标注小规模的数据集提供便利;然后利用深度神经网络模型对易标注图像预测标注的区域和类别,将预测区域的边界信息及标签类别写入json格式文件,实现自动标注,通过改进Yolact实例分割模型标注有重叠的图像;最后对自动标注得到的json文件进行人工校验调整。通过详细的自动标注结果对比试验(如小麦、葡萄、圣女果等),证明该智能标注系统在智慧农业领域表现出优异的性能;通过人工校正占比、标注时间、标注效率等定量对比实验,验证了相比传统人工标注方法具有更高的效率;通过与其他已有的智能图像标注系统对比实验,发现提出的智能图像标注系统更适用于智慧农业领域。这一研究结果为智慧农业图像标注提供了一种全新的智能化标注工具,具有重要的应用价值和推广前景。This study proposed an intelligent image annotation system for smart agriculture,aiming to solve the problem of time-consuming image annotation in smart agriculture scenes.Firstly,based on the idea of active learning,a novel query strategy was designed for image filtering.The images to be annotated were classified into difficult to annotate and easy to annotate images.The difficult to annotate images were used to optimize the automatic annotation model,which was beneficial for improving the accuracy of automatic annotation and providing convenience for annotating small-scale datasets;then,a deep neural network model was used to predict the annotated regions and categories of easily annotated images.The boundary information and label categories of the predicted regions were written into a json format file to achieve automatic annotation.By improving the Yoact instance segmentation model,overlapping images were annotated;finally,the json format file obtained from automatic annotation were manually verified and adjusted.Through detailed comparative experiments of automatic annotation results(such as wheat,grapes,Saint Mary's fruit,etc.),it wash proven that the intelligent annotation system showed excellent performance in the field of smart agriculture;through quantitative comparative experiments such as manual correction proportion,annotation time,and annotation efficiency,it was verified that compared to traditional manual annotation methods,it had higher efficiency;through comparative experiments with other existing intelligent image annotation systems,it was found that the proposed intelligent image annotation system was more suitable for the field of smart agriculture.This research result provides a new intelligent annotation tool for smart agricultural image annotation,which has important application value and promotion prospects.We believe that this intelligent image annotation system will have a profound impact on the future field of smart agriculture.

关 键 词：智慧农业 图像标注 主动学习 自动标注 深度学习 

分 类 号：S-3[农业科学]