Memorizing Structure-Texture Correspondence for Image Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Traditional autoencoder-based anomaly detection methods struggle with accurately identifying anomalies due to potential well-reconstructed abnormal images.
• Existing techniques often fail to distinguish subtle anomalies, limiting their sensitivity in critical applications.

Purpose of the Study:

• To develop a more sensitive image anomaly detection approach by leveraging the relationship between image structure and texture.
• To improve the accuracy of anomaly detection in both industrial and medical imaging domains.

Main Methods:

• Proposes a novel method that reconstructs images by learning structure-texture correspondence using a memory mechanism (STCM).
• Utilizes two types of complementary structures (semantic and low-level) extracted by separate networks.
• Fuses reconstructions from different structures using learned attention weights and incorporates structural consistency as an additional anomaly detection metric.

Main Results:

• The proposed method demonstrates higher sensitivity to anomalies compared to traditional autoencoder-based techniques.
• Achieves effective anomaly detection on both industrial inspection and medical image datasets.
• The structure-texture correspondence approach proves robust in identifying deviations from normal image patterns.

Conclusions:

• The novel structure-texture correspondence method significantly enhances image anomaly detection capabilities.
• This approach offers a promising solution for sensitive anomaly identification in diverse imaging applications.
• The integration of complementary structural information and learned attention improves reconstruction accuracy and anomaly detection performance.