Loading...
Computer vision for sustainable waste management: from contamination detection to zero-shot classification
Citations
Altmetric:
Date
2025-12
Abstract
The global rise in waste generation poses significant challenges for sustainable management, particularly in areas of recycling efficiency, contamination reduction, and operational logistics. This thesis investigates the application of advanced computer vision methods to address these challenges, with a focus on realistic industrial and municipal waste collection environments. Across multiple studies, we examine the potential of state-of-the-art object detection and segmentation models for two critical tasks: detecting overfilled bins during collection routes and identifying contamination in densely cluttered recyclable waste streams. These tasks are studied using proprietary datasets derived from real-world collection routes, reflecting the complex and cluttered nature of waste management systems.
Building on this foundation, we explore the use of Vision Transformers for contamination detection, highlighting their advantages in capturing global contextual information and their comparative performance against convolutional approaches. Finally, we extend this work by evaluating zero-shot learning through vision-language models such as CLIP and OWL-ViT, demonstrating their potential for scalable, annotation-free waste classification in dynamic environments where new materials frequently emerge.
Together, these studies provide an integrated perspective on the trade-offs between dataset quality, annotation effort, computational cost, and scalability. The findings establish the feasibility of deploying computer vision systems to automate key aspects of waste collection and recycling, ultimately contributing to reduced contamination, more efficient operations, and the advancement of sustainable circular economy practices.
Supervisor
Description
Publisher
University of Limerick
Citation
Collections
Files
Loading...
Mewada_2025_Computer.pdf
Adobe PDF, 13.57 MB