Self-Supervision-Augmented Deep Autoencoder for Unsupervised Visual Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep autoencoders (AEs) show promise in visual anomaly detection (VAD).
• AEs detect anomalies by larger reconstruction errors on abnormal samples.
• Existing AEs struggle with similar reconstruction errors for normal and abnormal data due to shared low-level features.

Purpose of the Study:

• To propose a self-supervised representation-augmented deep AE for unsupervised VAD.
• To enlarge the gap in anomaly scores between normal and abnormal samples.
• To improve the accuracy of VAD by learning high-level semantic features.

Main Methods:

• Introduced autoencoding transformation (AT) as a self-supervision task (transformation reconstruction).
• Input original and transformed images into the encoder for latent representations.
• Decoder reconstructs both the original image and the applied transformation.

Main Results:

• The proposed method effectively enlarges the gap of anomaly scores between normal and abnormal samples.
• The model utilizes both image and transformation reconstruction errors for anomaly detection.
• Achieved superior performance compared to state-of-the-art VAD methods in extensive experiments.

Conclusions:

• The proposed self-supervised representation-augmented deep AE advances unsupervised VAD.
• Autoencoding transformation (AT) facilitates learning high-level visual semantic features.
• The method demonstrates validity and advancement in visual anomaly detection.