Few-Shot Anomaly Detection with Defective Sample Augmentation
PhD. Seminar by: Atefeh Gilvari
Date: 23rd June 2026
Time: 1:00PM – 2:00PM
Location: Online – Microsoft
Abstract:
Electric vehicles (EVs) rely on highly efficient manufacturing processes that rarely produce defective parts, posing a challenge for training vision-based anomaly detection (AD) models due to limited abnormal samples. To address this, we propose FSOME++, an few-shot anomaly detection (AD) framework that integrating a small number of augmented abnormal samples during learning features. This leads to improved representation learning, with a clearer separation between normal and anomalous samples in the latent space. FSOME++ is evaluated on both the MVTec public AD benchmark and a private rotated EV Metal Triangle dataset. Experimental results show consistent improvements in evaluation metrics, including F1 score, Balanced Accuracy, and AVC across multiple backbone architectures, with ResNet-34 achieving the best trade-off between performance and stability.
PhD Doctoral Committee:
Internal Reader: Dr. Arunita Jaekel
Internal Reader: Dr. Boubakeur Boufama
External Reader: Dr. Ning Zhang
Advisor(s): Dr. Ziad Kobti, Narayan C. Kar