Waste Classification and Management Using Computer Vision

This technical report presents the development and benchmarking of deep learning models for image-based waste classification utilizing the TACO dataset. The study compares various architectures, including custom CNNs, ResNet34, and Vision Transformers (ViTs), focusing on their classification accuracy, inference performance, and deployment feasibility. The findings indicate that architectural enhancements significantly improve performance, with ViTs achieving the highest F1 score due to their ability to model spatial relationships effectively. The report highlights that class imbalance is not the primary limiting factor in classification accuracy; rather, the visual distinguishability of waste categories varies independently of sample size. All models meet real-time processing requirements, demonstrating sub-millisecond inference times suitable for cost-effective edge deployment. The work underscores the potential of computer vision technologies to enhance waste management systems, from centralized sorting facilities to smart bins, ultimately contributing to improved recycling rates and reduced landfill emissions.