Advancing Anomaly Detection in Industrial Systems: A Comparative Study of Autoencoder, CNN, and ResNet50

Anomaly detection is essential for industrial systems that function efficiently and produce high-quality products. The MVTec dataset, which consists of 15 item categories with various defect kinds, such as contaminated, broken large, and broken small, is used for this work to compare the effectiveness of Autoencoders, Convolutional Neural Networks (CNNs), and ResNet50 for identifying anomalies. This study focuses on the drawbacks of conventional approaches and emphasizes Vision Transformers’ potential as a strong substitute for dealing with inter-class variability and data imbalance. Key findings include the necessity of uniform normalization across object classes and the importance of generalized models for adaptability in diverse industrial settings. On comparing the results, it shows Resnet50 has higher accuracy of 94.49 percent, Convolutional Neural Network has an accuracy of 86 percent and Autoencoder has an accuracy of 70 percent. The findings show that ResNet50 performs better in supervised anomaly detection tasks, however, Vision Transformers have a lot of potential for scalability and flexibility.