Automatic Pixel-Level Crack Detection on Dam Surface Using Deep Convolutional Network

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Dam safety is critical for hydropower stations, necessitating effective crack detection methods.
• Current deep learning object detection methods often provide only classification and rough localization of cracks.
• Pixel-level crack detection offers more precise data for dam health assessment.

Purpose of the Study:

• To develop a deep convolution network for pixel-level crack detection on dam surfaces.
• To improve the accuracy and intuitiveness of crack detection for dam health assessment.

Main Methods:

• Utilized unmanned aerial vehicles (UAVs) for dam surface image acquisition.
• Created a labeled crack dataset by manually annotating cropped images.
• Designed and trained a novel Crack Detection on Dam Surface (CDDS) deep convolution network architecture.

Main Results:

• The CDDS network achieved recall of 80.45%, precision of 80.31%, F-measure of 79.16%, and IoU of 66.76% in crack segmentation.
• Compared performance against ResNet152-based, SegNet, UNet, and FCN models.
• Demonstrated superior performance of the CDDS network on the test dataset.

Conclusions:

• The proposed CDDS network effectively performs pixel-level crack detection on dam surfaces.
• The CDDS method provides more accurate and intuitive results compared to existing approaches.
• This technique enhances dam health assessment and ensures hydropower station safety.