Abstract:
"The concerns in the waste management domain are ever-increasing concerns that demand
innovative solutions. Adequate waste segregation is a crucial step in the waste management
process that can generate various issues if not done correctly. Although various experiments and
research are taking place to solve the problem of waste classification, significant gaps still need to
be bridged. Accordingly, through this study, the author discussed leveraging state-of-the-art
technologies to bridge some of the existing gaps. Throughout the study, the challenges of
identifying different types of visible waste that are not neatly separated, as well as the
impracticality of existing methods, are discussed.
The study uses the Yolov8 model, which utilizes advanced YOLO techniques to accurately detect
objects that overlap with each other. It also employs mechanisms such as center distance and data
augmentation techniques such as mixup, mosaic, and overlay augmentation. The author's goal is
to develop a mobile application that can perform real-time detection of waste in a live camera feed.
The application should have the ability to detect objects in complex environments such as
overlapping objects.
The performance metrics of the Yolov8 model, which has been trained for waste detection,
demonstrate its effectiveness in accurately distinguishing between various classes. With a mean
Average Precision (mAP) of 0.913 at an Intersection over Union (IoU) threshold of 0.5, the model
exhibits a strong capability in classifying and precisely localizing objects within images.
Moreover, the model achieves an optimal balance between precision and recall with an F1 score
of approximately 0.87 at a confidence threshold of about 0.495, indicating a robust classification
performance across different classes."
