Adaptive Valve Selection Algorithm for PET Bottle Sorting Equipment Based on Segment Anything Model

Polyethylene terephthalate bottle (PET) sorting equipment applies AI recognition technology to detect PET bottles on a conveyor belt, then transmits the category and location information to the lower-level control system, which operates the air valves to sort materials. For large or irregularly shaped items, the pressure of single air valve is insufficient, which means the manual assignment of multiple valves is required to handle the material. The existing methods often rely on the simple coordination of adjacent air valves around the center point of objects. If the activation point exceeds the actual boundary of objects, the surrounding items will be ejected with a certain risk, which results in the sorting accuracy reduced. This paper proposes an air valve selection algorithm based on the object detection combined with the keypoints location. The proposed model predicts the keypoints on the left and right boundaries of the object to determine the necessary air valves, while improving the sorting of entire bottles. Additionally, we propose an automatic keypoint labeling algorithm based on the segment anything model (SAM) by using the bounding boxes from object detection. Experimental results show that the proposed algorithm improves sorting accuracy without increasing manual labeling costs.