Detection and Evaluation of Anomaly Severity in Glassware Production Line Using Deep Neural Networks

Quality control of products in mass production is a critical challenge before market release. Traditional methods, such as human inspection, require skilled labor and are time-consuming and costly for high production volumes. In this study, artificial intelligence models were evaluated for anomaly detection using the MVTec Anomaly Detection dataset to enhance quality control processes. The focus was on the bottle subset, which includes images of both normal and defective bottles. The dataset was first organized and balanced to prevent bias in deep learning models. A custom neural network model was then designed and compared with VGG۱۹ and ResNet-۵۰ for classifying normal and defective bottles. Additionally, to simulate the quality control process more accurately, the U-Net model was employed for semantic segmentation to identify the precise locations of anomalies. The results indicate that data augmentation and proper model selection significantly improve anomaly detection accuracy. The proposed model performed effectively in product classification, while the U-Net model demonstrated strong performance in assessing anomaly severity.

کلیدواژه ها:

Anomaly detection ، deep learning ، product classification ، semantic segmentation