UTRAD

UTRAD applies a U-Transformer architecture with transformer-based autoencoders to detect and localize anomalies by reconstructing deep pre-trained feature tokens.

Key Features:

• Transformer-Based Autoencoders: Treats deep pre-trained features as dispersed word tokens within a transformer architecture to enable informative feature reconstruction and stabilize training.
• Feature Reconstruction over Raw Images: Focuses on reconstructing informative feature distributions rather than raw images to improve localization and training stability.
• Multi-Scale Pyramidal Hierarchy with Skip Connections: Integrates a multi-scale pyramidal hierarchy and skip connections to detect both structural and non-structural anomalies across scales.
• Decomposed Attention Layers: Decomposes attention into multi-level patches to reduce computational cost and memory usage relative to conventional transformer models.

Scientific Applications:

• Industrial Dataset (MVtec AD): Detects and localizes defects in industrial imagery using learned feature reconstruction.
• Retinal-OCT: Detects anomalies in retinal optical coherence tomography images.
• Brain-MRI: Identifies anomalies in brain magnetic resonance imaging scans.
• Head-CT: Locates abnormalities in head computed tomography scans.

Methodology:

Processes deep pre-trained features as tokens via transformer-based autoencoders, reconstructs informative feature distributions instead of raw images, integrates multi-scale pyramidal hierarchies with skip connections, and decomposes attention into multi-level patches to reduce computational cost and memory usage.