Threshold-Based BRISQUE-Assisted Deep Learning for Enhancing Crack Detection in Concrete Structures.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Automated visual inspection systems for concrete structures rely on accurate crack detection.
• Convolutional Neural Networks (CNNs) performance is hindered by low-quality images.
• Evaluating image dataset suitability is crucial for reliable deep learning models.

Purpose of the Study:

• To assess the impact of image degradations (Gaussian noise, blur) on crack detection using VGG16.
• To explore the correlation between image quality metrics (BRISQUE) and CNN performance.
• To propose a method for optimizing training and testing for reduced computational costs.

Main Methods:

• Utilized the BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator) method to quantify image degradations.
• Trained and evaluated the VGG16 model on datasets with varying levels of Gaussian noise and blur.
• Analyzed the relationship between BRISQUE scores and crack classification metrics (accuracy, F1 score, MCC).

Main Results:

• A strong correlation was found between lower BRISQUE scores and improved crack classification performance.
• Higher accuracy, F1 score, and Matthew's Correlation Coefficient (MCC) were achieved with less degraded images.
• Implementation of a BRISQUE score threshold can optimize model training and testing.

Conclusions:

• Image quality is a critical factor for the success of CNN-based crack detection.
• BRISQUE is a sensitive metric for evaluating image quality in the context of crack detection.
• Utilizing BRISQUE thresholds can lead to more efficient and cost-effective automated visual inspection systems for structural health monitoring.