Garbage Recognition based on YOLOv8-seg

An Instance Segmentation Approach for Automated Waste Sorting

Abstract

Rapid urbanization and population growth have led to a sharp increase in municipal solid waste, bringing severe challenges to environmental sustainability and public health. Effective waste classification and recycling are critical to alleviating these problems. However, traditional manual sorting is labor-intensive, inefficient, and prone to human error. In recent years, computer vision and deep learning have provided promising solutions for automated waste recognition. Object detection models can identify garbage and generate bounding boxes, but they struggle to capture the irregular shapes of waste, which is crucial for robotic grasping and automated sorting systems. To address this limitation, this paper explores the application of YOLOv8-seg, a state-of-the-art one-stage instance segmentation algorithm. It enables accurate garbage recognition and pixel-level segmentation in complex environments. We trained and evaluated the YOLOv8-seg model on the public TACO (Trash Annotations in Context) dataset. The model can simultaneously perform waste classification and generate instance masks. The network architecture-especially the C2f module and decoupled segmentation head-is analyzed in detail to clarify its feature extraction and mask generation mechanisms. Experimental results show that YOLOv8-seg achieves a strong balance between accuracy and inference speed, outperforming traditional two-stage models such as Mask R-CNN in real-time scenarios. Specifically, the model achieves a mask mAP@50 of 82.6% while maintaining a high frame rate suitable for edge deployment. This work provides a solid technical foundation for building intelligent waste-sorting robots and supports the development of automated environmental management.