DepthCrackNet: A Deep Learning Model for Automatic Pavement Crack Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Infrastructure Maintenance

Background:

• Manual pavement crack detection is labor-intensive and time-consuming, hindering efficient road safety management.
• Automated crack detection is crucial but faces challenges due to crack variability, diverse pavement materials, and surface anomalies.

Purpose of the Study:

• To propose an effective deep learning model, DepthCrackNet, for automated pavement crack detection.
• To enhance the accuracy and efficiency of crack segmentation in road infrastructure.

Main Methods:

• Developed DepthCrackNet, a U-Net shaped model incorporating a Double Convolution Encoder (DCE) for efficient feature extraction.
• Integrated the TriInput Multi-Head Spatial Attention (TMSA) module to capture diverse spatial relationships and contextual information.
• Utilized the Spatial Depth Enhancer (SDE) module to augment feature extraction capabilities for improved segmentation.

Main Results:

• DepthCrackNet achieved a mean Intersection over Union (mIoU) of 77.0% on the Crack500 dataset.
• The model attained an mIoU of 83.9% on the DeepCrack dataset, demonstrating high performance in crack detection.
• Experimental results validate the model's effectiveness in segmenting pavement cracks accurately.

Conclusions:

• DepthCrackNet offers a robust and efficient solution for automated pavement crack detection.
• The proposed model, with its novel modules, significantly advances the state-of-the-art in road surface analysis.
• This technology has the potential to enhance road safety and maintenance strategies through improved crack identification.