Lightweight concrete crack recognition model based on improved MobileNetV3

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Concrete crack detection is crucial for structural health monitoring.
• Existing lightweight models may lack accuracy and robustness in crack identification.
• Attention mechanisms offer potential for improving deep learning models in image analysis.

Purpose of the Study:

• To develop an improved lightweight concrete crack recognition model.
• To introduce the novel C//Sim attention mechanism by combining CA and SimAm attention.
• To evaluate the performance and robustness of the proposed model against existing methods.

Main Methods:

• The study proposed the C//Sim attention mechanism, a parallel connection of CA and SimAm attention mechanisms.
• MobileNetV3 was enhanced with the C//Sim attention mechanism to create the MobileNetV3-C//Sim model.
• The model was trained and validated on self-built and public datasets for crack detection.

Main Results:

• The MobileNetV3-C//Sim model demonstrated superior accuracy, recall, precision, and F1 scores compared to the original MobileNetV3.
• The parallel connection of attention mechanisms proved more effective than single mechanisms or concatenation.
• The C//Sim mechanism exhibited the smallest performance reduction under noise, indicating strong robustness and noise immunity.

Conclusions:

• The proposed C//Sim attention mechanism significantly enhances lightweight concrete crack detection performance.
• The MobileNetV3-C//Sim model offers improved accuracy, robustness, and parameter efficiency.
• This research provides a valuable reference for lightweight concrete crack identification using deep learning.