Recognition of Plastic Bottle Region Using Depth-Anything-V2 Incorporating the Monocular Depth Estimation Model

In recent years, the environmental burden caused by the disposal of plastic bottles has become a growing concern, making recycling increasingly necessary. However, many of the collected bottles still have caps attached or contain other contaminants, necessitating proper sorting. Currently, the sorting of plastic bottles from other waste in recycling plants is largely done manually. In Japan, however, the declining labor force caused by an aging society and low birth rate has exacerbated the labor shortage, creating an urgent need for automation of such tasks. To address this problem, automated sorting systems using robotic arms in conjunction with bottle recognition have been proposed. This study focuses on the recognition of plastic bottles. Specifically, we improve recognition performance by adding depth information to RGB images using a monocular depth estimation model, Depth-Anything-v2, and perform semantic segmentation with DeepLab v3+. Compared to conventional methods, the proposed approach improves both recognition accuracy and boundary clarity between the bottle body and cap.