Artificial intelligence based approaches for detection and assessment of building cracks

This paper addresses the key challenges faced by building crack detection in practical engineering applications, such as insufficient robustness in complex environments, high computational complexity of models, and the semantic gap from detection to quantitative assessment. It proposes an end-to-end crack detection and quantitative analysis method based on lightweight neural networks and multi-scale feature fusion (LWN-MFF). The method constructs a complete architecture including data preprocessing, feature extraction, feature fusion, and detection output. Specifically, an adaptive illumination compensation algorithm (AICA) is designed to improve the quality of input images. An improved ResNet-50 backbone network is adopted, introducing a detail-preserving branch and deformable convolution to enhance the extraction ability of crack geometric features. A cross-layer feature pyramid (CLFP) structure is further proposed, combining bidirectional feature calibration and feature importance reweighting mechanisms to achieve effective multi-scale feature fusion. Finally, a spatial attention mechanism is integrated in the detection head module to simultaneously complete crack localization and precise regression of morphological parameters. Experimental results on public datasets such as Concrete Crack Images for Classification and SDNET2018 show that this method maintains high accuracy while significantly reducing model complexity, with an AP@0.5 of 84.3%. The number of parameters is significantly reduced, and the computational cost is 8.9G FLOPs, which are reduced by 41% and 45% respectively compared to the mainstream lightweight model YOLOv11s. Ablation studies further confirm the efficacy of each innovative component and their synergistic effect. The integration of multi‑scale feature fusion and attention mechanism boosts crack localization detection accuracy by 9.5% points. This method achieves excellent lightweight characteristics while maintaining industry-leading accuracy, providing a breakthrough solution for real-time crack detection and quantitative analysis on embedded devices.