Generiranje pozitivnih učnih slik za nadzorovano večrazredno detekcijo površinskih anomalij

This thesis tackles the problem of limited defective samples in industrial surface anomaly detection and classification, by leveraging generative artificial intelligence for creating synthetic data. Manufacturing processes typically produce few faulty items, creating significant obstacles for conventional machine learning methods like supervised learning, that require substantial amounts of labeled training data to build effective detection and classification systems. We explore the use of the FLUX.1 Kontext diffusion model to generate synthetic yet realistic surface anomalies on industrial product images, specifically applied to the MVTec Hazelnut dataset. Our approach develops a comprehensive pipeline that produces both anomalous images and corresponding segmentation masks, enabling effective training of anomaly detection and anomaly classification models. Through systematic evaluation using established methodologies, we compare models trained with synthetic data augmentation against baseline model trained solely on original data. Results demonstrate improvements in anomaly detection, localization and classification capabilities, showing that generative models can effectively address the fundamental challenge of limited defective samples in industrial quality control.