Deep Convolutional Neural Network Optimization for Defect Detection in Fabric Inspection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Textile Manufacturing

Background:

• Traditional image processing struggles with complex backgrounds in fabric defect detection.
• Existing deep learning architectures are often over-parameterized for defect detection, leading to slow inference.
• Yolov4 shows limitations in detecting elongated fabric defects.

Purpose of the Study:

• To develop an optimized deep learning model for accurate and efficient fabric surface defect detection.
• To improve inference speed of deep learning models for fabric defect detection through network pruning.
• To address the limitations of current network architectures in handling fabric defect characteristics.

Main Methods:

• Proposed a novel approach combining network pruning with Bayesian optimization for automated tuning of pruning parameters.
• Retrained pruned networks and utilized a defect feature map prediction followed by an image processing pipeline for final defect judgment.
• Validated the method on two self-made and two public fabric defect datasets.

Main Results:

• Achieved significant reductions in inference time (up to 40.52% on GPU, 37.03% on embedded systems) with minimal drop in Intersection over Union (IoU) (max 2.15%).
• Post-image processing, the optimized model reached high detection accuracies across datasets (92.75% to 95.6%, and 81.82% on a challenging dataset).
• Demonstrated that Yolov4 is not optimal for detecting long and narrow fabric defects.

Conclusions:

• Network pruning guided by Bayesian optimization offers an effective strategy for accelerating deep learning-based fabric defect detection.
• The proposed method balances model efficiency and detection performance, making it suitable for real-world applications.
• The optimized approach provides a more efficient alternative to standard deep learning models for fabric defect identification.