Scholarlink

Anomaly detection is crucial in automated quality control systems across various manufacturing domains. This research presents a deep learning-based approach to detect anomalies in metal casting components using Convolutional Neural Networks (CNN), Visual Geometry Group 16 (VGG16), and MobileNet models. An annotated dataset of 7348 grayscale images representing Defective and Non-defective metal parts was used to train and test the models. Data preprocessing techniques were employed to improve generalization, and performance was assessed using multiple metrics such as Accuracy, Precision, Recall, F1-Score, Specificity, and AUC-ROC. All the models achieved values of more than 97% for every metric. In particular, MobileNet has got 99.62% and 99.99% in F1 score and AUC-ROC score, respectively, with a training time of 350.1 s only. In fact, MobileNet demonstrated a favorable balance between detection performance and computational efficiency, making it suitable for real-time deployment. Furthermore, the performance of our model outperformed the results available in the literature. The results confirm the potential of deep learning architectures in achieving reliable and automated anomaly detection in industrial settings.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Developing High Performance Anomaly Detection Models with Deep Learning Algorithms and Data Augmentation for Enhancing Visual Inspection of Metal Manufacturing

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rajeeb Das,
  • Dillip Rout

摘要

Anomaly detection is crucial in automated quality control systems across various manufacturing domains. This research presents a deep learning-based approach to detect anomalies in metal casting components using Convolutional Neural Networks (CNN), Visual Geometry Group 16 (VGG16), and MobileNet models. An annotated dataset of 7348 grayscale images representing Defective and Non-defective metal parts was used to train and test the models. Data preprocessing techniques were employed to improve generalization, and performance was assessed using multiple metrics such as Accuracy, Precision, Recall, F1-Score, Specificity, and AUC-ROC. All the models achieved values of more than 97% for every metric. In particular, MobileNet has got 99.62% and 99.99% in F1 score and AUC-ROC score, respectively, with a training time of 350.1 s only. In fact, MobileNet demonstrated a favorable balance between detection performance and computational efficiency, making it suitable for real-time deployment. Furthermore, the performance of our model outperformed the results available in the literature. The results confirm the potential of deep learning architectures in achieving reliable and automated anomaly detection in industrial settings.