计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (5): 160-165.DOI: 10.3778/j.issn.1002-8331.1811-0267

• 模式识别与人工智能 • 上一篇    下一篇

多线索植物种类识别

罗娟,蔡骋   

  1. 西北农林科技大学 信息工程学院,陕西 杨凌 712100
  • 出版日期:2020-03-01 发布日期:2020-03-06

Multi-Cue Plant Species Identification

LUO Juan, CAI Cheng   

  1. College of Information Engineering, University Northwest A&F, Yangling, Shaanxi 712100, China
  • Online:2020-03-01 Published:2020-03-06

摘要:

大多数关于自动植物识别的现有研究,集中于识别植物的单一器官,例如,花、叶或果实。使用单个器官的植物识别不够可靠,因为许多不同的植物却有着极其相似的器官。对于野外直接采集的图片,通常都有着复杂的背景,这也是目前的植物图像识别准确率不高的又一个原因。为了克服图像识别中的这两个难题,提出一种基于迁移学习的多线索植物识别方法,采用深度卷积神经网络(Deep Convolutional Neural Networks,DCNN)通过迁移学习,训练花、果、叶和整株的单器官分类器,根据各个分类器预测的标签和得分进行多器官融合识别。在PlantCLEF2017数据集上证明了模型有效性,并且植物识别性能得到了极大的提升。

关键词: 深度卷积神经网络, 迁移学习, 植物识别, 多线索

Abstract:

Most existing research on automated plant identification has focused on identifying single organs of plants, such as flowers, leaves or fruits. Plant identification using a single organ is not reliable because many different plants have very similar organs. For pictures taken directly in the wild, there are usually complex backgrounds, which is another reason why the accuracy of plant image recognition is not high. In order to overcome these two problems in image recognition, this paper proposes a multi-cue plant recognition. Specifically, a Deep Convolutional Neural Network(DCNN) is used to learn the single organ classifiers of flowers, fruits, leaves, and whole plants by transfer learning, and finally multi-organ fusion recognition is done based on the labels and scores predicted by each classifier. This paper demonstrates the efficiency of the model on the PlantCLEF2017 dataset, and the plant recognition performance has been greatly improved.

Key words: Deep Convolutional Neural Networks(DCNN), transfer learning, plant identification, multi-cue