Street waste segmentation and area classification using YOLOv8 and YOLOv11 with fuzzy inference

Computer Engineering and Informatics Department, Politeknik Negeri Bandung, 40559 Bandung, Indonesia

^* Corresponding author: trisgelar@polban.ac.id

Abstract

A deep learning-based solution for street waste analysis is presented, utilizing advanced instance segmentation models. To enhance the models’ resilience in from variations in real-world environments, a comparative analysis was performed between YOLOv8n and YOLOv11n, utilizing geometric and color data augmentation techniques. To transform the quantitative outputs of the models, specifically the segmented waste area and confidence scores, into a practical qualitative classification of waste density, such as Low, Medium, or High, a novel fuzzy inference system has been developed. The results suggest that YOLOv11n consistently outperformed YOLOv8n, achieving improved mAP50(M), a measure of segmentation accuracy, of 0.525. Furthermore, the effectiveness of both models was notably enhanced due to the incorporation of color augmentation. The fuzzy inference system offers a practical and transparent evaluation of waste accumulation. The results of our research provide a robust basis for the development of a cost-efficient, AI-driven system aimed at enhancing and overseeing municipal waste management practices.