Structural stability is essential for public safety, and early detection of building cracks is critical for maintenance. This study proposes Swin-Crack, an AI model based on the Swin-Transformer architecture, for effective crack detection. Preprocessing techniques such as Gaussian Blurring and Canny Edge Detection enhanced data quality and model learning. Swin-Crack outperforms CNN-based models, achieving 1.4% higher detection accuracy while remaining computationally efficient. It operates on standard mobile devices, enabling real-time detection and reducing maintenance costs. This study demonstrates the potential of integrating advanced preprocessing with AI to improve safety in construction, with future work focusing on predictive analytics for long-term structural monitoring.