Deep Learning Image-Based Fusion Approach for Identifying Multiple Apparent Diseases in Concrete Structure.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Traditional concrete disease detection methods struggle with pixel-level quantification and efficiency.
• Standalone object detection and semantic segmentation have limitations in identifying multiple concrete defects.

Purpose of the Study:

• To propose a deep learning fusion approach for identifying and quantifying typical visible diseases in concrete structures.
• To develop a fusion network combining YOLO and UNet models for rapid and accurate defect detection.

Main Methods:

• A deep learning fusion network integrating YOLO and UNet was developed.
• YOLO model filtered images with visible diseases; UNet segmented disease pixels.
• Quantification methods were proposed for disease measurements (length, width, area).

Main Results:

• The fusion network achieved a mean average precision of 0.72 and a Dice score of 0.82 on a dataset of 1488 images.
• The approach identified diseases approximately 50% faster than using UNet alone.
• Quantification showed satisfactory precision with relative errors below 11.07% for various defect types.

Conclusions:

• The proposed deep learning fusion network effectively identifies and quantifies concrete diseases.
• This method offers a robust and efficient solution for structural health monitoring of concrete infrastructure.
• The fusion approach significantly improves detection speed and quantification accuracy compared to standalone models.