Visual Anomaly Detection in Production Line : Metal Sleeve Dataset under Real Conditions
Pysyvä osoite
Kuvaus
Opinnäytetyö kokotekstinä PDF-muodossa.
Modern manufacturing increasingly relies on automated visual inspection to detect surface defects, yet real industrial environments remain difficult for anomaly detection systems. Production imagery rarely behaves like benchmark data: surfaces are reflective, lighting is uneven, defects can be subtle, and samples vary naturally from one another. This thesis examines deep learning approaches to visual anomaly detection in industrial surface inspection, with a particular focus on a refined SimpleNet pipe-line evaluated on a custom dataset of reflective steel sleeves. The study addresses a key limitation in current industrial anomaly detection research: many published results are obtained on controlled benchmark datasets such as MVTec AD, which do not fully represent the challenges of real production data. To examine this gap, the sleeve dataset was constructed in an MVTec-style format, with manually produced pixel-level masks for anomalous samples. Refined SimpleNet was trained on only defect-free images and evaluated alongside INP-Former, a transformer-based baseline. The work also includes hyperparameter tuning, augmentation experiments, qualitative error analysis, and a data leakage test. The results show that SimpleNet performs well on the MVTec AD categories, indicating that its architecture is dependable. However, the sleeve dataset performs worse, suggesting the challenge lies more in the data than in the model's architecture. INP-Former provides stronger defect localisation but weaker image-level classification. The augmentation study shows that strong geometric transformations improve image-level discrimination, while offline augmentation is more useful for localisation. The data leakage experiment further demonstrates how easily performance can be inflated on small custom datasets if the train-test split is not carefully controlled. Overall, the thesis shows that unsupervised anomaly detection can be useful for custom industrial inspection tasks, but only when dataset design, image capture conditions and evaluation discipline are treated as central parts of the method. The study concludes with practical recommendations for building new MVTec-style datasets for reflective metallic parts and highlights the need for larger, more realistic industrial datasets.