Building construction crack detection with BCCD YOLO enhanced feature fusion and attention mechanisms.

An effective algorithm for detecting cracks in bare concrete structures in building construction, capable of identifying small targets, is essential for safeguarding buildings. Unfortunately, detecting and precisely locating cracks in buildings is challenging due to variations in crack size, material texture, and inconsistent data quality. This study proposes a YOLO-based model for bare concrete crack detection, designated as BCCD-YOLO. Firstly, this model optimizes the Path Aggregation Network (PAN) by introducing lateral skips and weighted feature fusion mechanisms, improving the multi-scale fusion capability of bare concrete crack features. Secondly, a newly designed EC2f module is introduced into the neck network, adopting a channel-wise attention weighting mechanism to enhance feature fusion while maintaining computational efficiency. To further improve feature extraction, this study proposes an innovative SAC2f mechanism, which effectively integrates spatial and channel attention mechanisms for improved feature interaction. Empirical findings demonstrate that the refined model achieves a 3.3% increase in precision for detecting cracks in bare concrete structures within the gathered dataset. This contributes to the accurate and timely identification of structural cracks, thereby enhancing maintenance efficiency and preventing potential failures in building construction.