A swin transformer-based hybrid reconstruction discriminative network for image anomaly detection

Jin Jin ¹, Yuanping Xu ², Hui He ¹

⁰School of Software Engineering, Chengdu University of Information Technology, Chengdu, 610225, China.

Scientific Reports +

|

September 30, 2025

View abstract on PubMed

Summary

This study introduces a new network for industrial anomaly detection, improving the identification of small defects and performance in noisy settings. The Swin Transformer-Based Hybrid Reconstruction Discriminative Network (SRDAD) enhances accuracy and precision in defect detection.

Area Of Science

Computer Vision
Machine Learning
Industrial Automation

Background

Convolutional Neural Networks (CNNs) face challenges in detecting small anomalies and maintaining robustness in noisy industrial environments.
Existing methods often struggle with the complexities of real-world industrial defect detection.

Purpose Of The Study

To develop an advanced anomaly detection network for industrial applications.
To enhance the detection of small anomalies and improve performance in noisy conditions.

Main Methods

Proposed the Swin Transformer-Based Hybrid Reconstruction Discriminative Network (SRDAD).
Integrated a Swin-Unet for normal image reconstruction and a convolutional Unet for anomaly contrast discrimination.
Utilized hierarchical window attention and contrastive learning for improved detection and localization.

Main Results

SRDAD achieved competitive performance on the MVTec AD dataset.
Demonstrated improvements of 0.6% in detection accuracy and 0.7% in localization precision.
Showcased enhanced capability in detecting small anomalies and maintaining performance in noisy environments.

Conclusions

The SRDAD network effectively addresses limitations of traditional CNNs in industrial anomaly detection.
The hybrid reconstruction-discrimination approach offers superior performance for defect identification.
SRDAD shows significant potential for practical industrial defect detection applications.

Keywords:

Anomaly Detection Image Reconstruction Swin Transformer Unet