ADVANCING MANUFACTURING QUALITY CONTROL:YOLOV7-BASED BOTTLE DEFECT DETECTION WITH MULTI-SCALE TRAINING AND OPTIMIZATIONS

Abstract

Guaranteeing manufacturing quality control in bottle production often encounters incompetence and inaccuracies due to reliance on manual or conventional methods for defect detection. This study addresses these challenges by leveraging the YOLOv7 object detection framework, recognized for its real-time processing capabilities and accuracy. Using the Roboflow platform, we organized and augmented a dataset focusing on bottle defects, such as missing caps and deformed surfaces, and fine-tuned a model pretrained on the COCO dataset for our specific application. Multi-scale training and optimizations, including anchor box refinement and spatial attention mechanisms, enhanced the model's detection performance across varying defect types and conditions. The proposed approach achieves a mean Average Precision (mAP) of 83%, with a precision of 86.7% and a recall of 95%, reflecting robust performance. The study highlights the feasibility of real-time defect detection in industrial environments, reducing production waste and ensuring highquality standards. Practical deployment considerations, combined with a focus on real-time processing, underscore the model's potential to replace traditional methods, paving the way for automated quality control solutions. These findings contribute to the advancement of computer vision in industrial applications, setting a precedent for future research in automated defect detection.