Multi-Stage Computer Vision Framework with Ensemble Learning for Real-Time Glass Packaging Defect Detection in Industrial Applications

Jonah Alfred Mekel
Rick Resa Wahani
Motulo, Firmansyah Reskal
Alfred Noufie Mekel
Tineke Saroinsong
Tammy Tinny V. Pangow
Jerry Heisye Purnama
Jedithjah Naapia Tamedi Papia

Frontier Advances in Applied Science and Engineering
Vol. 3 No. 2 (2025)

Transparent glass packaging inspection presents significant challenges for automated quality control systems due to optical complexities including reflections, refractions, and low-contrast defect patterns. This research develops a comprehensive multi-stage computer vision framework integrating specialized algorithmic modules with ensemble machine learning for real-time defect detection in industrial glass packaging lines. The framework implements four specialized detection stages: (1) meniscus-corrected liquid level measurement using dual-camera validation and polynomial surface fitting, (2) seal integrity assessment through Circular Hough Transform combined with geometric, texture, and color feature extraction, (3) lid positioning evaluation via calibrated geometric centroid analysis with tolerance-based classification, and (4) multi-method contamination detection integrating color aberration analysis, histogram-based particle detection, and morphological operations. The system employs an ensemble classification architecture combining modified MobileNetV2 convolutional neural network with Random Forest classifier, optimized for edge computing deployment. Industrial validation at PT AQUWAR Bintang Semesta demonstrated 91.6% overall detection accuracy with 347 milliseconds average processing time per container across 2,847 test samples spanning multiple defect categories. The modular framework architecture enables independent optimization of detection stages while maintaining real-time processing capabilities, providing a robust foundation for transparent packaging quality control in high-volume manufacturing environments.