A multi-scale attention mechanism for detecting defects in leather fabrics

Area of Science:

• Industrial Engineering
• Computer Vision
• Materials Science

Background:

• Defect detection is crucial for quality control in leather production.
• Challenges include diverse defect sizes, locations, and subtle inter-class variations.

Purpose of the Study:

• To develop an advanced automated method for accurate leather defect detection.
• To enhance the identification of small and similar defect types.

Main Methods:

• Proposed a Multi-Layer Residual Convolutional Attention (MLRCA) approach to improve feature representation.
• Integrated MLRCA into a Feature Pyramid Network (FPN) creating ML-FPN for multi-scale fusion.
• Implemented a Side-Aware Boundary Localization (SABL) detection head for precise defect positioning.

Main Results:

• The MLRCA-based model achieved high performance metrics: AP (83.4), AP50 (89.7), AP75 (85.6).
• Excellent detection of small (AP_S: 71.3), medium (AP_M: 89.9), and large (AP_L: 88.9) defects.
• Demonstrated superior capability in distinguishing between similar defect categories.

Conclusions:

• The proposed MLRCA and ML-FPN approach offers a feasible and effective solution for automated leather defect detection.
• Provides new insights for enhancing industrial quality control in leather manufacturing.
• The method shows significant potential for detecting subtle and minor defects.