A Crack Segmentation Model Combining Morphological Network and Multiple Loss Mechanism.

Area of Science:

• Engineering
• Computer Vision
• Deep Learning

Background:

• Traditional crack detection methods are costly, inefficient, and inaccurate.
• Image-based crack detection is crucial for structural health monitoring in pipelines, pavements, and dams.

Purpose of the Study:

• To propose an effective crack segmentation network for improved accuracy and efficiency.
• To address challenges in crack detection, including varying resolutions and illumination conditions.

Main Methods:

• Utilized U-Net for multi-scale feature extraction to identify cracks of different resolutions.
• Applied white-top hat and black-bottom hat transforms for morphological processing to mitigate illumination effects.
• Implemented a multi-loss mechanism to enhance segmentation accuracy across different scales.

Main Results:

• Achieved average performance metrics: ODS 75.7%, OIS 73.9%, AIU 36.4%, sODS 52.4%, and sOIS 52.2% on five datasets.
• Demonstrated superior crack segmentation performance compared to state-of-the-art methods.
• Ablation experiments confirmed the effectiveness of individual algorithmic modules.

Conclusions:

• The proposed crack segmentation network significantly improves detection accuracy and efficiency.
• The combination of U-Net, morphological transforms, and multi-loss mechanisms offers a robust solution for crack detection.
• This method provides a valuable tool for structural health monitoring applications.