<![CDATA[Waste management has become a pressing global challenge due to rapid urbanization and population growth, leading to increased environmental pollution and resource depletion. Automating waste segregation using deep learning can significantly improve recycling efficiency and reduce manual labor. In this study, five deep learning models: YOLOv8, YOLOv5, MobileNetV2, Single Shot Multibox Detector (SSD), and Google’s pre-trained EfficientNet, were evaluated for real-time waste detection and classification. A dataset consisting of 15,150 images across 12 waste categories was used for training and testing, representing common household waste materials and evaluated using real-world objects such as water bottles, glass bottles, cans, and used face masks. Testing was conducted on both a high-performance workstation and a Raspberry Pi 4 edge device to assess detection accuracy, inference speed, and practical deployment feasibility. The results indicate that YOLOv8 achieved the highest average accuracy of 83% with reliable performance across diverse object types, whereas EfficientNet achieved moderate accuracy of approximately 60% with higher inference latency. Lightweight models such as MobileNetV2 and SSD were computationally efficient but exhibited lower accuracy, at 42% and 41%, respectively, particularly when handling irregular or overlapping objects. These findings demonstrate that YOLOv8 provides the most effective balance between accuracy and real-time performance, making it well suited for intelligent waste sorting systems.]]>
Copyrights © 2026
