Boundary-Aware Axial Attention Network for High-Quality Pavement Crack Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Transportation Systems

Background:

• Pavement crack detection is crucial for road maintenance in intelligent transportation systems.
• Existing methods struggle with diverse environmental factors, complex topologies, and intensity variations, leading to poor generalization.
• Severe foreground-background imbalance and overfitting issues hinder the performance of current crack detection models.

Purpose of the Study:

• To develop an innovative approach for high-quality pavement crack detection.
• To address the limitations of existing methods in handling diverse crack conditions and data imbalance.
• To improve the generalization and performance of pavement crack detection models.

Main Methods:

• Proposed the Boundary-Aware Axial Attention Network (BAAN), featuring a hierarchical encoder-decoder architecture.
• Incorporated position-guided axial attention (PAA) modules for capturing precise spatial structures and contextual information.
• Introduced a boundary regularization module (BRM) to enhance foreground-background discrimination and a boundary refinement loss (BRL) to manage data imbalance.

Main Results:

• The BAAN demonstrated consistent outperformance against state-of-the-art methods across four crack datasets.
• The proposed method effectively handles diverse crack conditions, including variations in illumination and intensity.
• BAAN achieved superior performance while requiring fewer computational resources compared to existing approaches.

Conclusions:

• The Boundary-Aware Axial Attention Network (BAAN) offers a significant advancement in pavement crack detection.
• BAAN effectively overcomes challenges related to data imbalance and diverse environmental factors.
• The proposed method presents a computationally efficient and high-performing solution for intelligent road maintenance.