Next-Gen Coffee Bean Inspection System Achieves 97.4% Precision with Dual-Sided AI – CoffeeTalk

A groundbreaking integrated mechatronic and computer vision system is poised to revolutionize green coffee bean defect inspection, offering unprecedented accuracy and real-time insights for coffee industry professionals. The new system, detailed in a recent research paper, addresses the long-standing challenge of comprehensive bean inspection by automating dual-sided presentation and leveraging advanced AI.

The innovative device tackles the dual-sided inspection challenge head-on with a compact mechatronic apparatus that sequentially presents individual green coffee beans to an imaging system under controlled LED illumination. This design is crucial for detecting defects often missed by single-sided inspection methods.

Mean per-class detection accuracy rose from 0.727 to 0.908, a +25.7% relative increase, directly attributable to the mechatronic design.

At the heart of the system’s intelligence is its computer vision module, which utilizes the YOLOv8 AI architecture. After rigorous evaluation against VGG-16, VGG-19, and ResNet-50, YOLOv8 emerged as the top performer, achieving an impressive 97.4% precision and 96.5% mAPIoU=0.5 in defect detection.

The research also highlights the system’s practical deployment capabilities, noting that YOLOv8n (the nano variant) can operate efficiently on Raspberry Pi-class hardware without a GPU. This makes the technology accessible for embedded grading devices, with an inference time of approximately 470 ms per image.

To train the robust AI model, a comprehensive labeled image dataset was generated, comprising 9600 images across 12 defect classes. This dataset was enriched through systematic data augmentation strategies, specifically addressing the intrinsically long-tailed nature of defect datasets where rare defects often carry the highest commercial penalties.

One of the system’s most compelling features for industry professionals is its real-time “track” mode implementation on a Raspberry Pi 4. This functionality simultaneously classifies defects, counts beans by category, and monitors throughput, providing immediate actionable information for production optimization.

The development process was structured around four integrated sections: mechatronic system design and construction, electronics and control architecture, dataset generation and augmentation, and AI model training and performance evaluation. This holistic approach ensures a robust and reliable inspection solution.

This advanced inspection system promises to enhance quality control, reduce waste, and improve operational efficiency across the coffee supply chain, offering a significant leap forward for coffee industry professionals seeking to optimize their processes.

Read More @ MDPI