Testing Methodology for a Computer Vision System for Glass Container Quality Control on a Robotic Conveyor

Abstract

This paper presents a comprehensive testing methodology for a prototype cascaded computer vision system deployed on a robotic conveyor mockup for automated quality control of glass containers. The work extends previously proposed approaches to synthetic training data generation by shifting the focus to empirical validation of deep learning algorithms on real product samples using low‑power edge platforms. The physical experimental setup is implemented on a laboratory belt conveyor equipped with a Raspberry Pi 5 single‑board computer, a Hailo‑8L NPU accelerator, a Raspberry Pi AI Camera (IMX500) sensor and a programmable LED ring light. A real COCO‑compatible dataset with key classes of industrial defects is collected and a structured test protocol is developed. The system is evaluated in terms of detection accuracy, inference performance, thermal stability under long‑term load, and robustness to motion blur and illumination changes. Experimental results confirm that the proposed architecture achieves industrial‑level inspection performance on edge devices: mAP reaches 0.95, the false positive rate does not exceed 5%, and the processing speed consistently exceeds 30 FPS. This demonstrates the feasibility of integrating such hardware–software complexes directly into real‑time production conveyor lines.