MBE-YOLOv8: enhancing building crack detection with an advanced YOLOv8 framework

Abstract Buildings, over prolonged periods, are susceptible to developing various types of cracks, which are often small and exhibit low contrast, leading to challenges in accurate detection. Missed detections and false positives due to these characteristics can result in delayed repairs, thereby compromising structural integrity and safety. Therefore, real-time detection of building cracks is essential to maintain the longevity and safety of infrastructures. In response to these challenges, we present an optimized version of the YOLOv8 model, referred to as MBE-YOLOv8, designed specifically for building crack detection. The core enhancement involves restructuring the backbone of YOLOv8 with the integration of the multi-dimensional collaborative attention mechanism, significantly improving feature interrelationships and the extraction capabilities of the backbone network. Additionally, we introduced a Weighted Feature Fusion Network (BiFPN) and developed a novel BiFPN-L structure to enhance feature fusion and detection accuracy, particularly for small targets. The efficient channel attention (ECA) mechanism was also incorporated into the model’s neck, leading to the design of a new EC2f structure that improves the model’s adaptability to scale variations and overall feature extraction efficiency. A comparative analysis with the original YOLOv8 model demonstrated that MBE-YOLOv8 achieved performance improvements with P, R , and mAP@0.5 values of 78.6%, 67.0%, and 73.4%, respectively. These figures represent increases of 4.8, 3.8, and 4.1 percentage points compared to the previous version of the YOLOv8 model. This advancement has significantly bolstered the capability to detect cracks in buildings. Furthermore, the enhanced model preserves a compact size of 3.0 M while sustaining a high frame rate (FPS), rendering it highly deployable for applications related to crack detection.