Deep Learning and Continual Learning Techniques for Plant Image Analysis
The research presented in this thesis addresses the application of deep learning on digital images, particularly plant images. The exponential growth of publicly available image datasets, mainly due to the wide accessibility of smartphones and digital cameras, has sparked a surge in deep learning re...
Saved in:
| Main Author: | |
|---|---|
| Contributors: | |
| Format: | Doctoral Thesis |
| Language: | English |
| Published: |
Philipps-Universität Marburg
2024
|
| Subjects: | |
| Online Access: | PDF Full Text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The research presented in this thesis addresses the application of deep learning on digital images, particularly plant images. The exponential growth of publicly available image datasets, mainly due to the wide accessibility of smartphones and digital cameras, has sparked a surge in deep learning research across various domains. Online platforms like iNaturalist, GBIF, and Zooniverse offer hundreds of millions of images, including digitized herbarium scans from museums and collections worldwide. These serve as invaluable resources for ecological and biodiversity research. While plant images from natural environments can provide an excellent resource for studying species distributions and ecological traits, herbarium scans offer additional advantages, such as analysis of visual and structural plant features in a standardized format relevant for analyzing phenological traits of species spanning hundreds of years. This thesis presents innovative methods for species recognition, trait extraction, and plant organ detection by leveraging novel deep learning techniques for image recognition and object detection. While recognizing the successful implementation of these approaches, the thesis also highlights crucial challenges such as data imbalance and limited availability of labeled datasets. The thesis addresses these challenges and proposes an innovative, data-free continual learning approach for training a model on continuously arriving data while also mitigating data imbalance. This approach enables the integration of new data of unknown distribution into existing models while preserving the previously learned knowledge without access to the prior data. Through a combination of practical deep learning applications and theoretical insights, the research presented in this thesis contributes significantly to advancements in ecological research and continual learning. |
|---|---|
| DOI: | 10.17192/z2024.0111 |