PCTC-Net: A Crack Segmentation Network with Parallel Dual Encoder Network Fusing Pre-Conv-Based Transformers and...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Structural Health Monitoring

Background:

• Cracks are prevalent defects in structures, necessitating manual inspection for maintenance.
• Manual crack detection is labor-intensive, costly, and inefficient for large-scale applications.
• Automated crack detection using computational resources is an active area of research.

Purpose of the Study:

• To develop an efficient and accurate crack segmentation model that overcomes the data-intensive nature of transformers.
• To improve crack detection performance despite limited availability of fine-grained crack datasets.
• To propose a novel network architecture that effectively fuses convolutional neural networks and transformers.

Main Methods:

• Proposed a parallel dual encoder network, PCTC-Net, integrating Pre-Convolution (Pre-Conv) based Transformers and Convolutional Neural Networks (CNNs).
• Introduced a Pre-Conv module to optimize color channels before transformer input, mitigating data requirements.
• Evaluated PCTC-Net on benchmark datasets: DeepCrack, Crack500, and Crackseg9k.

Main Results:

• PCTC-Net demonstrated superior generalization performance compared to the state-of-the-art DTrC-Net.
• The proposed model achieved higher stability and improved F1 scores in crack segmentation tasks.
• Experimental results validated the effectiveness of the PCTC-Net architecture in addressing data limitations.

Conclusions:

• PCTC-Net offers an effective solution for crack segmentation, particularly in scenarios with limited labeled data.
• The fusion of CNNs and transformers with the Pre-Conv module enhances model efficiency and accuracy.
• The findings contribute to advancing automated structural health monitoring and maintenance technologies.