Automated Real-Time Bottle Defect Detection Using YOLOv8, BoT-SORT Tracking, and Audio Alerts

Abstract: The quality assurance in the manufacturing process requires precise and accurate identification of defects in the product in real-time in order to avoid operational losses. Traditionally, the inspection process in a manufacturing environment relies on human observation or classical image processing techniques. These methods are more likely to be erroneous and time-consuming and cannot adapt to changing defect characteristics without major programming changes. In this context, the development of an automated bottle defect detection system using the YOLOv8 framework, BoT- SORT multi-object tracking algorithm, and audio alerts in real-time is proposed in this research. A dataset with five classes of bottle defects: cap, label, crumbled, no-cap, and not-crumbled, with a total of 1,250 samples, has been created and fine-tuned with the pre-trained YOLOv8n model with 80 epochs. A script test2.py has been written in order to run the model in real-time on both images and videos, drawing bounding boxes with their respective probabilities. Finally, the performance of the model has been evaluated with a score of 0.965 mAP@0.5, 0.953 macro F1, and 47 FPS on GPU.

Keywords: YOLOv8; Bottle Defect Detection; Machine Vision; BoT-SORT Tracking; Object Detection; Industrial Automation; Transfer Learning; Deep Learning; OpenCV; Ultralytics