A diffusion-based framework with transformer blocks for anomaly detection

doi:10.1016/j.asoc.2026.115477

A diffusion-based framework with transformer blocks for anomaly detection

2026-05-19

0

PRE

AI

DOI：10.1016/j.asoc.2026.115477

Original

Original request for help

Save

Abstract

En 中文

• Transformer-guided Diffusion Framework. We propose DTAD, a Transformer-based diffusion framework for industrial anomaly detection, which replaces the traditional UNet backbone with a UViT architecture to better capture global contextual dependencies in visual feature reconstruction. • Feature-Guided Modulation Mechanism. A novel Feature-Guided Modulation module is introduced to enhance the representation alignment between normal and anomalous regions by injecting semantic priors into the diffusion denoising process. • Comprehensive Performance Improvement. Extensive experiments on MVTecAD and VisA datasets demonstrate that DTAD achieves superior performance in both image-level and pixel-level anomaly detection compared with existing diffusion-based methods, including DiffusionAD, DDAD, and DiAD.

Keywords:

Transformer

Diffusion Model

Anomaly Detection

Feature-Guided Modulation

UViT

A diffusion-based framework with transformer blocks for anomaly detection

Abstract

Journal

Organization