FewShot-CrackDet an Improved LoRA-Based Few-Shot Intelligent Detection Model for Bridge and Building Structural Cracks

Cracks in bridge and building structures are critical safety hazards. Traditional detection relies on manual work or data-hungry AI models, suffering from low efficiency and high costs. This paper proposes FewShot-CrackDet, an improved LoRA-based few-shot detection model. A pre-trained ViT serves as the backbone, enhanced by a cross-layer feature fusion LoRA module to reduce overfitting in few-shot scenarios. A multi-scenario dataset with 400 annotated samples (bridge/building cracks, harsh environments) is constructed. Experiments show the model achieves mAP@ 0.5 of 89.3%, 93.7%, 96.2% under 5/10/50-shot scenarios, outperforming baseline models by 6.8%-12.5% with 75% fewer parameters. It reduces engineering data collection costs and provides an efficient solution for structural safety monitoring.