Object Detection with Scarce Data: Scans of Entomological Boxes using Active Learning
Files
Lowenthal_88781900_Usoltsev_42361800_2025.pdf
Open access - Adobe PDF
- 19.73 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- Insects are essential to ecosystem stability, yet global warming threatens their populations worldwide. Monitoring insect biodiversity is critical but challenging, especially using physical entomological collections. Digitizing these collections and applying AI for automatic insect detection can revolutionize ecological studies. This thesis focuses on memory and data-efficient AI methods for detecting and counting insects in digitized entomological box images. Leveraging a custom dataset of 139 images and limited hardware (8 GB RAM MacBook Air), three approaches were developed: (1) post-detection filtering of pre-trained models, (2) fine-tuning lightweight YOLOv8 models via a dynamic multi-phase method, and (3) active learning strategies to reduce required training data. Notably, AL enabled achieving comparable detection performance to full dataset training while using 50% less data. Our best model sets a new state-of-the-art in insect detection, reaching an map@50 of 0.799, while significantly lowering computational complexity, memory usage and hence overall electricity consumption. This work advances sustainable AI applications in biodiversity monitoring and contributes foundational tools for large-scale digitized insect collections.