Camera-Only Unsupervised Visual Anomaly Detection for IEQ-Oriented Building Operations

Maintaining indoor environmental quality (IEQ) requires the timely identification of surface defects such as structural cracks and water stains which are often overlooked by conventional environmental sensors. However, the application of automated computer vision in the built environment has been hindered by the scarcity of labeled defect data and the impracticality of training supervised models for every facility. To bridge this gap, this study introduces a pioneering framework that adapts the WinCLIP paradigm to indoor building operations. This represents the first successful application of a CLIP-based zero-shot anomaly detection model to the domain of indoor surface monitoring. By leveraging the generalization power of pre-trained vision-language models, the proposed approach eliminates the need for defect-specific training or fine-tuning. The framework is validated on four distinct datasets comprised of a large-scale primary benchmark for comprehensive pixel-level and image-level evaluation and three supplementary datasets for assessing generalization capability. The findings confirm that foundational vision models can effectively distinguish defective surfaces from normative ones across diverse indoor settings. While localization precision varies by defect morphology, the system provides reliable visual guidance to prioritize maintenance actions. Ultimately, this work demonstrates that training-free anomaly detection is a privacy-preserving and scalable solution. It opens a new pathway for integrating advanced AI into facility management without the burden of laborintensive data annotation.