A Novel Computer-Vision Approach Assisted by 2D-Wavelet Transform and Locality Sensitive Discriminant Analysis for...

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Accurate detection of concrete cracks is crucial for structural health monitoring.
• Traditional methods and existing deep learning models often face challenges with computational cost and performance in adverse conditions.

Purpose of the Study:

• To introduce FastCrackNet, a computationally efficient deep learning model for detecting tiny concrete cracks.
• To evaluate FastCrackNet's performance against established Convolutional Neural Network (CNN) architectures in challenging imaging scenarios.

Main Methods:

• FastCrackNet utilizes a fully connected network combined with 2D-wavelet image transform and Locality Sensitive Discriminant Analysis (LSDA) for feature extraction and reduction.
• The model was tested on concrete images with noise and shadows, comparing its speed and accuracy against GoogleNet and Xception.

Main Results:

• FastCrackNet demonstrated superior speed and accuracy compared to GoogleNet and Xception.
• The model achieved over 60x and 80x higher classification efficiency than GoogleNet and Xception, respectively.
• FastCrackNet proved robust and stable across various input image sizes and batch sizes.

Conclusions:

• FastCrackNet provides a computationally efficient and highly accurate solution for concrete crack detection.
• The proposed method establishes new performance benchmarks for image classification in difficult environmental conditions.
• FastCrackNet offers a reliable and resilient alternative for structural health monitoring applications.