Synthetic Anomaly Generation

Unsupervised anomaly detection models in Anomalib rely on synthetic anomalies to calibrate decision thresholds, but the current Perlin-based pipeline produces unrealistic and static anomalies, leading to unstable threshold estimation and poor generalization to real defects. This project proposes a hybrid, on-the-fly synthetic anomaly generation pipeline that combines multi-scale fractal masks with self-sourced Poisson blending to generate realistic, localized anomalies with controlled difficulty levels. By generating diverse anomalies dynamically during validation, the model can learn a stable and reliable decision threshold without requiring labeled anomaly data. The project will deliver a structured literature survey, integration of the proposed pipeline into Anomalib, and a benchmarking study comparing it with the current baseline using PyTorch and OpenVINO, focusing on threshold stability, detection performance, and CPU efficiency.