Directional Lighting-Based Deep Learning Models for Crack and Spalling Classification.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Autonomous inspections of concrete structures require effective external lighting, especially in low-light conditions.
• Uniformly diffused lighting has limitations in detecting complex crack patterns in concrete.
• Existing methods struggle with accurate concrete defect classification under challenging illumination.

Purpose of the Study:

• To propose novel algorithms using directional lighting for improved concrete defect classification.
• To address the limitations of uniformly diffused lighting in crack detection.
• To enhance the accuracy and reliability of autonomous concrete inspections.

Main Methods:

• Developed two algorithms: a fused neural network and a multi-channel neural network.
• Fused neural network utilizes maximum intensity pixel-level image fusion from directional images.
• Multi-channel neural network creates a five-channel image representing different lighting directions (Right, Down, Left, Up, Diffused).

Main Results:

• The multi-channel neural network model demonstrated superior performance.
• Achieved high evaluation metrics: 96.6% accuracy, 96.3% precision, 97% recall, and 96.6% F1 score.
• Outperformed the FusedNet and other existing literature models without increasing evaluation time.

Conclusions:

• Directional lighting significantly improves concrete crack classification accuracy.
• The multi-channel neural network is a promising approach for autonomous concrete inspections in low-light environments.
• Future work will explore extending these techniques to white-box methods.