Pavement Cracks Segmentation Algorithm Based on Conditional Generative Adversarial Network.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Road cracks pose significant financial and safety risks.
• Accurate segmentation of pavement cracks is challenging due to complex backgrounds and various disturbances.
• Existing methods struggle with the precise detection of cracks in real-world conditions.

Purpose of the Study:

• To propose an advanced pavement crack segmentation method using a conditional generative adversarial network (cGAN).
• To enhance the accuracy and robustness of crack detection on road surfaces.
• To address the limitations of current techniques in handling complex backgrounds and noise.

Main Methods:

• Utilized a U-net3+ architecture with an attention module in the generator to highlight crack features and suppress noise.
• Employed a PatchGAN discriminator, trained with original images, manual annotations, and generated segmentations.
• Developed a weighted hybrid loss function to optimize the difference between generated and annotated crack images.

Main Results:

• The proposed cGAN method demonstrated superior performance in segmenting pavement cracks.
• The attention module effectively fused channel and spatial features for more accurate crack identification.
• The method successfully eliminated disturbances, achieving high accuracy on the Crack500 dataset.

Conclusions:

• The developed conditional generative adversarial network offers a robust solution for pavement crack segmentation.
• The integration of attention mechanisms and a novel loss function significantly improves detection accuracy.
• This technique provides a reliable tool for pavement maintenance and safety assessment.