Pixel-Wise Crack Detection Using Deep Local Pattern Predictor for Robot Application.

Area of Science:

• Engineering
• Computer Science
• Materials Science

Background:

• Manual inspection of concrete bridges for cracks is time-consuming and costly.
• Ensuring bridge safety necessitates efficient and accurate crack detection methods.

Purpose of the Study:

• To develop a robust, pixel-wise crack detection algorithm for concrete bridges.
• To address challenges in crack detection, including environmental noise and achieving high accuracy.

Main Methods:

• A local pattern predictor (LPP) was developed using convolutional neural networks (CNNs) for feature extraction.
• Each pixel was classified as crack or non-crack using LPP with contextual patches.
• CNN output (confidence map) was post-processed to delineate crack areas.

Main Results:

• The proposed algorithm demonstrated robust performance in detecting cracks from real concrete bridge images.
• Experimental results validated the effectiveness of the CNN-based approach.

Conclusions:

• The developed algorithm offers an efficient and accurate solution for robotic vision-based crack detection in concrete bridges.
• This method enhances the preservation and safety of critical infrastructure.