Convolutional autoencoder based on latent subspace projection for anomaly detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Image anomaly detection (AD) faces challenges with high-dimensional, noisy, and imbalanced data.
• Existing unsupervised deep learning methods struggle with discriminative feature learning in single latent spaces.

Purpose of the Study:

• To propose a novel autoencoder framework, LSP-CAE, for robust unsupervised image anomaly detection.
• To address limitations of single latent spaces by employing mutually orthogonal complementary subspaces.

Main Methods:

• Introduced Latent Subspace Projection (LSP) mechanism within an autoencoder.
• Developed two trainable, complementary latent subspaces: Latent Image Subspace (LIS) for normal features and Latent Kernel Subspace (LKS) for irrelevant information.
• Validated using fully-connected networks on real-world medical datasets.

Main Results:

• LSP-CAE demonstrated superior performance in anomaly detection across four public datasets.
• The method effectively enhances learning of diverse features by separating normal and irrelevant information.
• Anomaly scores derived from projection norms in the two subspaces accurately identified anomalies.

Conclusions:

• LSP-CAE offers a powerful and generalizable approach for unsupervised image anomaly detection.
• The complementary subspace strategy significantly improves feature discriminability compared to single latent spaces.
• The proposed method achieves state-of-the-art results, particularly for complex and imbalanced image datasets.