Riassunto analitico
Visual Anomaly detection(VAD) plays a crucial role in industrial quality control, where identifying defects with minimal supervision is essential. This thesis explores the integration of few-shot learning techniques into an VAD pipeline, leveraging a general-purpose segmenter to enhance robustness. The study evaluates two distinct anomaly detection paradigms: reconstruction-based models and student-teacher architectures, both commonly used in industrial scenarios. A common challenge in anomaly detection arises when anomalies occur not on the reference object but in the background, leading to potential misclassifications. To address this, we incorporate a few-shot general-purpose segmenter within the pipeline to improve object isolation. The segmenter is based on PANet, a few-shot segmentation model trained on the COCO 2017 dataset, enabling the system to generalize across diverse objects with limited annotated data. Segmentation is performed directly on the images for which an anomaly map is to be generated. Once the segmentation mask is obtained, it is used to filter out background regions from the anomaly map, ensuring that only anomalies on the segmented object are considered while disregarding extraneous detections in the background. Experimental results demonstrate that incorporating this segmentation step enhances model robustness across various environments and improves generalization to unseen anomalies. These findings contribute to the development of more reliable and adaptive anomaly detection systems in industrial settings.
|