Earlier cracks identification is very crucial in structural building maintenance as it is the main signifier of building deterioration. Manual inspection processes are slow, expensive and can be easily compromised by human error. Though the You Only Look Once version 8 (YOLOv8) has emerged as a powerful framework for automated crack detection, it faces limitations in accurately detecting small, irregularly shaped, and partially obscured cracks due to feature loss in deeper network layers and insufficient pixel-level precision. This study addresses these limitations by strategically integrating five attention mechanisms into YOLOv8's architecture: Convolutional Block Attention Module (CBAM), Efficient Channel Attention (ECA), Selective Kernel Attention (SKA), Shuffle Attention (SA), and Global Attention Mechanism (GAM). The attention modules were placed at critical positions within the backbone and neck regions to enhance feature representation without compromising computational efficiency. Using a comprehensive dataset of 13,169 building crack images with 19,386 annotations, each attention-enhanced variant was trained and evaluated against the baseline YOLOv8 model. Results demonstrate consistent improvements across all attention mechanisms. CBAM achieved the highest segmentation accuracy with mask mean Average Precision (mAP) @0.5 of 0.820 (0.4% improvement), while ECA provided the most parameter-efficient enhancement, improving box precision by 3.5% with only 41 additional parameters. SKA excelled in recall performance, achieving 0.724 (1.0% improvement), valuable for comprehensive building crack detection. All variants maintained practical deployment feasibility, supporting real-time building inspection applications. The findings confirm that attention mechanism integration offers an effective approach to enhance YOLOv8 for building crack detection, providing empirical evidence for attention module selection based on specific deployment constraints and supporting the development of more reliable automated building inspection systems.
Owoeye et al. (Wed,) studied this question.