Pixel Intensity Resemblance Measurement and Deep Learning Based Computer Vision Model for Crack Detection and...

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image noise from drone capture and varied lighting degrades crack detection accuracy.
• Accurate crack severity classification is crucial for timely infrastructure maintenance and accident prevention.

Purpose of the Study:

• To develop and evaluate a novel technique for improving the efficiency and accuracy of crack detection in noisy images.
• To classify cracks based on severity levels for effective risk analysis and alerting.

Main Methods:

• A pixel-intensity resemblance measurement (PIRM) rule was proposed to classify noisy and noiseless images.
• Median filtering was employed for noise reduction.
• Crack detection utilized deep learning models: VGG-16, ResNet-50, InceptionResNet-V2, and Xception.
• A crack risk-analysis algorithm segregated images based on detected crack severity.

Main Results:

• The PIRM rule improved crack detection efficiency by up to 10% across different models (VGG-16, ResNet-50, InceptionResNet-V2, Xception).
• High accuracies were achieved for specific noise types: ResNet-50 (95.6% for Gaussian), InceptionResNet-v2 (99.65% for Poisson), and Xception (99.95% for speckle).
• The system provides alerts based on crack severity, enabling prompt action.

Conclusions:

• The proposed PIRM rule effectively enhances crack detection in noisy images, outperforming existing methods.
• The integration of deep learning models with PIRM and risk analysis offers a robust solution for infrastructure monitoring.
• This technique contributes to improved safety by enabling proactive maintenance through accurate crack assessment.