Attention-guided hybrid learning for accurate defect classification in manufacturing environments.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Industrial defect classification is challenging due to visual complexity and diverse defect types.
• Existing methods often struggle with multi-class classification across different object categories.

Purpose of the Study:

• To develop a robust hybrid deep learning framework for unified multi-class industrial defect classification.
• To improve accuracy and generalization capabilities in defect detection systems.

Main Methods:

• Integration of YOLOv11 for spatial features and EfficientNet-B7 for fine-grained representations.
• Incorporation of Convolutional Block Attention Module (CBAM) and Feature Pyramid Network (FPN) for attention-guided multi-scale refinement.
• Evaluation on MVTec-FS and a proprietary Window dataset.

Main Results:

• Achieved state-of-the-art accuracy: 91.90% on MVTec-FS and 96.13% on the Window dataset.
• Outperformed existing Convolutional Neural Network (CNN), transformer, and ensemble baselines.
• Ablation studies confirmed the performance contribution of each module.

Conclusions:

• The proposed hybrid deep learning framework demonstrates superior performance in industrial defect classification.
• The model exhibits strong generalization capabilities across different domains and defect types.
• The framework offers practical utility for industrial quality control.