CrackNet: A Hybrid Model for Crack Segmentation with Dynamic Loss Function.

Area of Science:

• Computer Science
• Civil Engineering
• Artificial Intelligence

Background:

• Infrastructure integrity is crucial for public safety, with cracks representing a common and significant form of damage.
• Visual-based automatic crack detection using deep learning faces challenges like complex crack patterns, background noise, and data imbalance.

Purpose of the Study:

• To develop an advanced deep learning model, CrackNet, for robust and accurate automatic crack detection in infrastructure.
• To address the limitations of existing methods by integrating Convolutional Neural Networks (CNNs) and transformers.

Main Methods:

• Proposed CrackNet, a hybrid network combining CNNs for local feature extraction and transformers for global dependency modeling.
• Introduced a strip pooling module to enhance segmentation of narrow cracks and suppress irrelevant background noise.
• Implemented an attention-based skip connection and mixed up-sampling in the decoder for detailed information restoration.
• Developed a joint learning loss function with dynamic weighting to handle severe class imbalance.

Main Results:

• CrackNet demonstrated superior performance compared to several established deep neural networks on three public crack datasets.
• The model achieved a particularly significant improvement in recall rate, indicating better detection of actual cracks.
• Experimental results validated the effectiveness of the hybrid architecture and proposed modules.

Conclusions:

• CrackNet offers a promising solution for automated crack detection in infrastructure, enhancing safety and maintenance.
• The hybrid CNN-transformer approach combined with specialized modules effectively addresses key challenges in visual crack analysis.
• The proposed joint learning loss is crucial for mitigating class imbalance issues in crack detection datasets.