Adjacent Image Augmentation and Its Framework for Self-Supervised Learning in Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep neural networks are advancing anomaly detection.
• Training requires normal and anomalous data, often causing class imbalance due to scarce anomaly data.
• Traditional augmentation methods struggle to preserve correlations between anomalies and their context.

Purpose of the Study:

• To introduce an adjacent augmentation technique for generating synthetic anomaly images.
• To preserve object shapes while distorting contours for enhanced correlation.
• To improve anomaly detection performance and mitigate class imbalance.

Main Methods:

• Adjacent augmentation generates synthetic anomaly images by preserving object shapes and distorting contours.
• Synthetic normal images are generated to learn detailed normal data features.
• A framework pairs training images with synthetic normal (positive) and anomaly (negative) images within batches.

Main Results:

• Adjacent augmentation captures high-quality anomaly features, outperforming existing methods in AU-ROC and AU-PR scores.
• The technique produces synthetic normal images, reducing sensitivity to minor variations.
• Achieved perfect AU-ROC and AU-PR scores (100%) using ResNet50 on the MVTec-AD bottle dataset.

Conclusions:

• Adjacent augmentation effectively compensates for limited anomalous features and mitigates class imbalance.
• The method enhances the ability to distinguish between normal and anomalous features.
• Further research is exploring the impact of anomalous pattern size on detection performance.