Abstract:
Traditional visual inspection of production lines is tedious, error-prone, and
costly, and therefore inappropriate for high-volume production environments. The aim
of this project is to investigate the potential of employing computer vision and
Convolutional Neural Networks (CNNs) to produce automatic fault detection on
manufacturing lines.
Since I was the AI engineer in charge of this project, model training, backend
development, model testing, accuracy improvement, hyperparameter tuning, fine tune
operations, and dataset expansion were activities that I was tasked with. The adopted
strategy was large-scale testing of a series of architectures like Anomalib framework
implementations, CNN models, HRNet, YOLO variants, and other deep learning
methods. All the models were thoroughly evaluated on performance metrics including
accuracy, precision, recall, F1-score, mean Average Precision (mAP), and processing
speeds.
Key results indicate notable performance improvements in defect detection.
Precise models achieved high accuracy with precision and recall maximized metrics
with processing rates compatible with real-time industrial operations. The mAP scores
and F1-scores validate consistent performance across different classes of defects and
categories of products. Findings confirm the technical feasibility of implementing
computer vision-based quality assurance systems in factories with the potential to
transform factory quality assurance by eliminating human mistakes and saving huge
sums of money through auto-defect detection.
