Transfer learning between crop types for semantic segmentation of crops versus weeds in precision agriculture
Transfer learning between crop types for semantic segmentation of crops versus weeds in precision agriculture
复制标题
DOI：
10.1002/rob.21869
复制
发表时间：
2020-01-01
影响因子：
8.3
通讯作者：
Cielniak, Grzegorz
中科院分区：
文献类型：
--
作者：
Bosilj, Petra;Aptoula, Erchan;Cielniak, Grzegorz
农业机器人依靠语义分割来区分作物和杂草，以进行选择性处理，提高产量和作物健康，同时减少化学品的使用量。深度学习方法最近实现了出色的分类性能和实时执行。然而，这些技术也依赖于大量的训练数据，需要大量的标记工作，这两者在精准农业中都是稀缺的。在不同的环境条件和作物生长阶段下，需要额外的设计努力来实现商业上可行的性能水平。在本文中，我们探讨了不同作物类型的基于深度学习的分类器之间的知识转移的作用，目的是减少新作物所需的再训练时间和标记工作。我们研究了具有不同作物类型和包含各种杂草的三个数据集的分类性能，并比较了使用像素级标记的数据与通过更耗时的注释分割输出过程获得的部分标记数据时所需的性能和再训练工作。我们证明了不同作物类型之间的迁移学习是可能的，并将训练时间减少了80%。此外，我们表明，即使用于再训练的数据是不完美的注释，分类性能是在2%的网络训练与辛苦注释像素精度数据。
Agricultural robots rely on semantic segmentation for distinguishing between crops and weeds to perform selective treatments and increase yield and crop health while reducing the amount of chemicals used. Deep-learning approaches have recently achieved both excellent classification performance and real-time execution. However, these techniques also rely on a large amount of training data, requiring a substantial labeling effort, both of which are scarce in precision agriculture. Additional design efforts are required to achieve commercially viable performance levels under varying environmental conditions and crop growth stages. In this paper, we explore the role of knowledge transfer between deep-learning-based classifiers for different crop types, with the goal of reducing the retraining time and labeling efforts required for a new crop. We examine the classification performance on three datasets with different crop types and containing a variety of weeds and compare the performance and retraining efforts required when using data labeled at pixel level with partially labeled data obtained through a less time-consuming procedure of annotating the segmentation output. We show that transfer learning between different crop types is possible and reduces training times for up to 80%. Furthermore, we show that even when the data used for retraining are imperfectly annotated, the classification performance is within 2% of that of networks trained with laboriously annotated pixel-precision data.