Bringing Attention to Image Anomaly Detection

Detecting anomalies in images is a task with several relevant real-world applications, e.g. industrial inspection. Building on the existing RIAD (Reconstruction by Inpainting for visual Anomaly Detection) framework, we introduce an attention-based component to improve the model performance. Furthermore we propose a different approach to image masking which leverages the selection of multiple random patches at a single scale in the original images. Through the provided experimental results we show how the novelties introduced by this work consistently improve the performance of the baseline approach over the various classes of the heterogeneous MVTec benchmark dataset across all the metrics considered.