Learning with fewer labels in deep learning for plant phenotyping

Chen, Feng (2022) Learning with fewer labels in deep learning for plant phenotyping. PhD thesis, University of Nottingham.

[img] PDF (Corrected Version of Thesis) (Thesis - as examined) - Repository staff only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution.
Download (9MB)


Deep learning has become an important approach in the research of plant phenotyping. However, training a highly-performing deep learning model generally requires a large amount of labeled data. These training data are normally labeled by human annotators and the labeling process is expensive and time-consuming. In plant phenotyping, precisely labeling the images often needs professional knowledge, for example, knowing the species of a plant or the appearance of an organ, disease or pest, which further increases the costs of annotating the data. Hence, exploring effective training methods that use less labeled data to achieve comparable performance can significantly boost the development of deep learning in plant phenotyping.

In this thesis, three categories of approaches are explored, namely active learning, semi-supervised learning, and weakly-supervised learning, with a common purpose -- reducing the dependence on labeled data. Novel approaches of learning with fewer labels are proposed and evaluated on various plant-phenotyping-related and generic tasks. In addition to boosting the performance of deep learning with fewer labels, the proposed methods are adaptive or require fewer parameters, so that the effort of optimizing the learning models is also reduced. These approaches are also evaluated on other computer vision tasks to show their generality.

In particular, we perform flower classification to demonstrate that active learning can reduce 30\% training data without sacrificing much performance. We perform semi-supervised aphid counting and crowd counting using Mean Teacher and FixMatch, which are two milestones in semi-supervised learning. Specifically, we improve the property and reduce the hyper-parameters of Mean Teacher by using an adaptive consistency weight. FixMatch is improved and adapted to heatmap regression using Monte-Carlo Dropout and a novel data augmentation method -- Shadowout. Finally, a scenario combining semi and weakly-supervised learning -- learning with incomplete pixel-level labels is explored on wheat spikelet localisation and crowd counting. We proposed to use a simple yet effective method -- an asymmetric loss function to learn well-performed models in this scene.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: French, Andrew
Tony, Tony
Subjects: Q Science > QA Mathematics > QA 75 Electronic computers. Computer science
Faculties/Schools: UK Campuses > Faculty of Science > School of Computer Science
Item ID: 71783
Depositing User: Chen, Feng
Date Deposited: 15 Dec 2022 09:33
Last Modified: 15 Dec 2022 09:33
URI: https://eprints.nottingham.ac.uk/id/eprint/71783

Actions (Archive Staff Only)

Edit View Edit View