Loading...
Towards Accurate Waste Recognition and Segmentation in Complex Environments
Ali, Muhammad
Ali, Muhammad
Files
Author
Supervisor
Department
Machine Learning
Embargo End Date
Type
Dissertation
Date
2024
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This thesis addresses the global waste crisis by leveraging advanced computer vision and deep learning techniques to improve waste classification, detection, and segmentation in complex environments, aiming to replace inefficient manual sorting methods. Key contributions include FlexPooling and CLIP-Decoder for accurate single-label and zero-shot multi-label classification, novel image enhancement techniques for underwater waste detection, and three segmentation architectures FreezeConnect, FANet, and COSNet that tackle challenges like background clutter and occlusions for pixel-level precision. Additionally, the Waste-Bench benchmarking framework evaluates Vision-Language Models (VLMs) in cluttered waste environments, identifying gaps and fostering the development of robust solutions. While limitations like computational efficiency remain, the findings contribute to advancing automated waste management systems and promoting environmental sustainability.
Citation
M. Ali, "Towards Accurate Waste Recognition and Segmentation in Complex Environments", PhD Dissertation, Machine Learning, MBZUAI, Abu Dhabi, UAE, 2024