Multi-Class Concrete Defect Classification Using Guided Semantic-Spatial Fusion and Squeeze-Excitation Enhanced...

Area of Science:

• Civil Engineering
• Computer Science
• Materials Science

Background:

• Concrete structures are prone to defects impacting safety and maintenance.
• Accurate defect detection and quantification are essential for structural health monitoring.
• Existing deep learning methods often lack multi-class defect identification capabilities.

Purpose of the Study:

• To develop an advanced deep learning framework for multi-class concrete defect detection and localization.
• To enhance the accuracy and reliability of automated concrete defect classification.
• To create a practical tool for non-destructive measurement of concrete defects.

Main Methods:

• A novel dataset of 2029 concrete defect images across five categories was compiled.
• The DenseNet201 model was enhanced with a guided semantic-spatial fusion module and squeeze-and-excitation architecture.
• Attention mechanisms were integrated to improve feature representation and defect region tracking.

Main Results:

• The proposed framework achieved a 5.6% accuracy improvement over the original DenseNet201 model.
• Experimental validation demonstrated the model's superiority in multi-class defect identification.
• The developed model effectively detects and localizes various concrete defects.

Conclusions:

• The enhanced deep learning framework offers superior performance for multi-class concrete defect detection.
• The study provides a reliable method for non-destructive measurement and classification of concrete defects.
• Integration into a graphical user interface facilitates practical application in structural maintenance.