Automatic Fabric Defect Detection with a Multi-Scale Convolutional Denoising Autoencoder Network Model.

Area of Science:

• Textile Manufacturing
• Computer Vision
• Machine Learning

Background:

• Manual fabric inspection is inefficient and imprecise for industrial quality control.
• Automated defect detection is crucial for improving textile manufacturing processes.

Purpose of the Study:

• To develop an unsupervised learning-based automated approach for fabric defect detection and localization.
• To overcome limitations of traditional manual inspection methods.

Main Methods:

• Utilized a convolutional denoising autoencoder network for image patch reconstruction at multiple Gaussian pyramid levels.
• Employed reconstruction residuals as indicators for pixel-wise defect prediction.
• Integrated detection results from multiple resolution channels for robust analysis.

Main Results:

• The method achieves high precision and acceptable recall rates in fabric defect detection.
• Demonstrated robustness across various textile fabric types, from simple to complex.
• Successfully detected and localized defects without manual intervention.

Conclusions:

• The proposed unsupervised learning method offers an efficient and accurate alternative to manual fabric inspection.
• The approach requires only defect-free samples for training, making it practical for real-world applications.
• Multi-modal integration enhances the robustness and accuracy of the automated fabric inspection system.