Texture conditioned GAN based self supervised framework for fabric defect detection.

Area of Science:

• Computer Vision
• Machine Learning
• Textile Manufacturing

Background:

• Automated fabric inspection is crucial for consistent textile quality in high-speed manufacturing.
• Existing methods often require extensive labeled data and struggle with complex textures, leading to inaccuracies.
• There is a need for data-efficient anomaly detection that preserves fine-grained texture semantics.

Purpose of the Study:

• To develop a novel, annotation-free framework for fabric defect detection.
• To address the limitations of supervised methods and improve detection on complex fabrics.
• To enhance the reliability of automated quality inspection systems.

Main Methods:

• Introduced the Self-Supervised Texture-Decomposition and Defect Amplification Network (STDAN).
• Decomposed fabric images into texture bases and anomaly cues using a texture-consistency encoder.
• Employed a generative defect amplification module and a contrastive self-supervised objective for feature-level learning.

Main Results:

• STDAN achieved superior detection accuracy and localization precision on public benchmarks.
• Demonstrated effectiveness in detecting small, low-contrast, and previously unseen defects.
• Showcased strong robustness to texture variations and illumination changes.

Conclusions:

• STDAN offers a scalable, annotation-free solution for reliable fabric quality inspection.
• The framework successfully combines texture decomposition and feature-level defect amplification.
• The method is practically applicable in industrial environments, improving defect detection capabilities.