A lightweight YOLOE visual-prompt framework for real-time multi-class waste detection

Abstract

Inefficient waste segregation remains a significant environmental challenge, particularly in regions where manual sorting processes are slow and pose serious health risks to workers. Previously developed automated waste classification systems often rely on high-cost hardware and cloud-dependent architectures. Moreover, these models struggle to handle material variability in real-world scenarios. Many conventional deep learning models also fail to effectively recognize irregular or poorly textured waste materials, which limits their practical applicability in real-world smart waste management environments. These operational limitations motivate the need for a lightweight, accurate, and easily deployable solution. In this study, YOLOE was selected due to its visual-prompt-based detection mechanism, which enables the incorporation of material-specific prompts and enhances recognition accuracy.

After getting the motivation of these challenges, this thesis work develops a real-time waste detection system which can detect four key waste categories such as glass, metal, plastic and bio-waste while using the YOLOE visual prompt object detection model. For the system deployment was done on a Raspberry Pi 5 which provides a low-cost, portable and efficient edge-based platform for the suitable waste management applications. The YOLOE model was introduced with visual prompts to get better results on material specific understanding and evaluate on static and real-time video streams. There were optimized preprocessing, detection, and post-processing procedures which was to ensure stable frame-wise inference on the real-time embedded device.

Experimental results indicate that the system of the following work achieves high-confidence detection rate for the glass, metal, and plastic objects whereas the bio-waste produces lower confidence rate compared to others due to its irregular structure and inconsistent texture patterns. During real time testing the Raspberry Pi 5 demonstrated smooth and more accurate performance with the continuous detection update and zero noticeable latency. In conclusion, this study shows the validity of the YOLOE visual-prompt model which can be effectively deployed for the edge-based real-time environmental waste classification.