Smartphone-based deep learning for real-time quality assessment in glass recycling
摘要
In modern recycling facilities, the precise separation of glass fractions is crucial for maintaining high-quality cullet streams and reducing contaminants. This paper presents a novel computer vision-based system for monitoring and analyzing glass recycling streams, focusing on flint, green and amber glass. Using smartphone-based image acquisition under both normal and ultraviolet (UV) illumination, samples are captured to detect and classify relevant contaminants, including lead-containing glass that fluoresces under UV light. The images are stored on a MinIO server, while associated metadata is managed in a MongoDB database. Object detection is performed with a YOLOv8-based model, and instance segmentation is refined using a Segment Anything-based module, classifying fragments into predefined categories such as Flint glass, Amber glass, Green glass, Blue glass, Lead glass under UV, Heat-resistant glass, Ceramics and porcelain, Stones, Metal wires and Metal springs. Detected objects are annotated, and their mass is estimated by applying gram-per-pixel calibration derived from prior measurements. The results are aggregated and transferred into a SQL database and further communicated to a programmable logic controller (PLC) via an OPC UA interface, enabling real-time monitoring and automated alarm generation if threshold values are exceeded. This approach provides a scalable, cost-efficient, and automated solution for enhancing glass recycling quality control.
A patent application related to the described procedure and device has been filed under the title “Verfahren und Vorrichtung zur Analyse eines Objektgemischs” (application no. A55101/2025).