Development and Evaluation of an OAK-D-Based Waste Classification System Using Transfer Learning
Supervisor Name
Emad Natsheh
Supervisor Email
e.natsheh@najah.edu
University
An-Najah National University
Research field
Artificial Intelligence
Bio
Dr. Emad Natsheh is an Associate Professor in the Faculty of Engineering and Information Technology at An-Najah National University. He holds a PhD in Artificial Intelligence and Renewable Energy and an MSc in Computer and Network Technology from Manchester Metropolitan University (UK), and a BSc in Computer Engineering from An-Najah National University. His research focuses on artificial intelligence, machine learning, smart systems, and renewable energy applications. Dr. Natsheh has published several papers in international journals and conferences and has experience in training in data science, Python programming, and data analysis.
Description
Effective waste segregation supports recycling efficiency, resource recovery, and the reduction of contamination between waste streams. Reliable automated classification is therefore important, particularly when systems must operate under varying environmental conditions with limited computational resources. This project aims to improve and evaluate an existing embedded waste classification system using an OAK-D camera and a pretrained EcoVision model based on MobileNetV3. The initial system classifies waste into four categories: metal, plastic, paper, and other. The proposed research addresses the practical challenge of maintaining accurate and consistent classification in the actual operating environment. Images will be collected and annotated from this environment and used to fine-tune the model. Object presence detection, background isolation, tracking, and prediction aggregation across multiple frames will also be refined to improve classification reliability. The research contribution lies in the systematic integration and experimental evaluation of these improvements within an embedded waste classification pipeline, providing evidence of their combined effectiveness under realistic deployment conditions. The improved system will be compared with the initial system using a separate test dataset. Evaluation will measure classification accuracy, per-class precision, recall, F1-score, and inference speed, with additional testing under different lighting and background conditions. The final model will be exported to ONNX and converted to Blob format for deployment on the camera’s Myriad X processor. The expected outcome is an experimentally validated system with improved classification reliability, together with practical findings that can guide the deployment of embedded vision systems for waste segregation.