AdapSyn: Anomaly detection based on triplet training with adaptive anomaly synthesis

Abstract

In the field of anomaly detection (AD), Few-Shot Anomaly Detection (FSAD) has gained significant attention in recent years. The goal of anomaly detection is to identify defects at the image level and localize them at the pixel level. Including defect data in training helps to improve model performance, and FSAD methods require a large amount of data to achieve better results. However, defect data is often difficult to obtain in experiments and applications. This paper proposes a more realistic method for simulating anomaly data, incorporating the synthesized anomaly data into the training process and applying it to the FSAD domain through multi-class mixed training. The anomaly data generated using our synthesis method closely resembles real anomalies. The corresponding anomaly synthesis method synthesizes the anomaly data from normal samples for the adaptively selected polygonal mesh region. We enhance the model’s ability to distinguish between positive and negative samples by incorporating synthesized anomaly data and normal data as triplets during training. This results in that more detailed features for normal samples will be noticed. During the testing phase, we obtain the feature distribution of normal images for a few unknown class normal samples to quickly adapt to the detection task of new categories. The effectiveness of the anomaly synthesis method was validated through experiments. Comparisons with advanced methods in the FSAD domain demonstrated that our method achieved competitive performance.