Prototypical Learning Guided Context-Aware Segmentation Network for Few-Shot Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Few-shot anomaly detection (FSAD) faces challenges due to the domain gap between pre-trained features and target scenarios.
• Limited normal samples in FSAD hinder the effectiveness of traditional anomaly detection methods.

Purpose of the Study:

• To propose a novel network, PCSNet, that addresses the domain gap in FSAD.
• To enhance feature descriptiveness and improve FSAD performance using limited normal samples.

Main Methods:

• Proposes a prototypical learning-guided context-aware segmentation network (PCSNet).
• Introduces a prototypical feature adaptation (PFA) subnetwork for feature compactness and anomaly separation.
• Incorporates a pixel-level disparity classification (PDC) loss for distinguishing subtle anomalies.
• Utilizes a context-aware segmentation (CAS) subnetwork with pseudo anomalies for pixel-level localization.

Main Results:

• PCSNet achieved superior FSAD performance on MVTec AD and MPDD datasets.
• Demonstrated high image-level area under the receiver operating characteristics (AUROCs) of 94.9% and 80.2% in an eight-shot scenario.
• Showcased promising results in real-world automotive plastic part inspection with limited training data.

Conclusions:

• PCSNet effectively bridges the domain gap in FSAD.
• The proposed method significantly enhances feature descriptiveness and FSAD performance.
• PCSNet offers a viable solution for anomaly detection with limited data in real-world applications.