Design of Machine Vision Detection System for Impurities in Liquid Bottles

This paper proposes an intelligent detection system for impurities in liquids based on machine vision. The system focuses on visual information acquisition and target recognition for impurities such as hair, scrap, and grain. In the experiments, lighting systems are compared to reduce irrelevant information, and the optimized top lighting is chosen to effectively capture impurity images and create a dataset of 3000 images. The improved visual enhancement (VE)-YOLOv8 deep learning detection algorithm incorporates a lightweight efficient channel attention (ECA) mechanism into the model's neck and integrates the Swin Transformer with a deformable attention mechanism into the backbone module to enhance the network's feature extraction capabilities and improve detection accuracy. In the optical comparison experiments, the analyses showed mean average precision (mAP 50) values of 87.3%, 91.5%, 97.1%, and 91.5% for the common, back, top, and bottom conditions, respectively. The mAP 50 and mAP 50–90 of VE-YOLOv8 are improved by 1.4% and 5.6% compared with the original YOLOv8 in optimized top lighting conditions. This research combines the optical lighting technique with the target detection algorithm to enhance the identifiable accuracy of impurities in liquid bottles.