Masked Transformer for Image Anomaly Localization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Image anomaly detection is crucial for applications like industrial inspection and medical imaging.
• Current deep learning methods often use image reconstruction, which can fail when anomalies are similar to normal data.
• Limitations exist in reconstruction-based anomaly detection when anomalies share characteristics with normal data.

Purpose of the Study:

• To develop a more robust image anomaly detection model.
• To overcome the limitations of traditional reconstruction-based deep learning approaches.
• To improve the accuracy and reliability of identifying visual anomalies.

Main Methods:

• A novel Vision Transformer architecture incorporating patch masking was developed.
• Input images are divided into patches, with each patch reconstructed solely from surrounding contextual information.
• Multi-resolution patches and their combined embeddings were utilized to enhance performance.

Main Results:

• The proposed patch masking approach demonstrated superior performance compared to traditional reconstruction methods.
• Utilizing multi-resolution patches significantly improved anomaly detection accuracy.
• The model achieved competitive results on benchmark datasets like MVTec and head CT.

Conclusions:

• The Vision Transformer with patch masking offers a promising alternative for image anomaly detection.
• Reconstructing patches from surrounding context effectively handles anomalous regions.
• Multi-resolution patch embeddings are key to enhancing the model's detection capabilities.