Low-Shot Unsupervised Visual Anomaly Detection via Sparse Feature Representation.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Manufacturing

Background:

• Visual anomaly detection is crucial for industrial manufacturing quality control.
• Existing methods based on pairwise similarity distance have limitations in generalization, especially with limited data.
• Absolute similarity distance struggles to extend comparisons beyond available samples, hindering performance in low-shot scenarios.

Purpose of the Study:

• To present a novel sparse feature representation anomaly detection (SFRAD) framework.
• To address the generalization challenges in anomaly detection, particularly in low-shot scenarios.
• To propose a new anomaly score metric using orthogonal matching pursuit (ASOMP).

Main Methods:

• Formulated anomaly detection as a sparse feature representation problem.
• Employed the orthogonal matching pursuit (OMP) algorithm for sparse representation.
• Introduced a basis feature sampling (BFS) algorithm for efficient memory bank construction, balancing coverage and representation.

Main Results:

• The SFRAD framework demonstrated superior performance across multiple benchmark datasets (MVTec AD, KolektorSDD, MNIST, CIFAR-10, etc.).
• Achieved state-of-the-art results in unsupervised anomaly detection.
• Significantly improved performance in low-shot anomaly detection scenarios.

Conclusions:

• SFRAD effectively combines absolute similarity and linear representation advantages.
• The proposed framework enhances generalization capabilities, especially for limited sample sizes.
• SFRAD represents a significant advancement in unsupervised and low-shot visual anomaly detection.