Thermal anomaly detection in walls via CNN-based segmentation

Abstract IRT (Infrared Thermography) is a commonly used non-destructive testing method for detecting thermal anomalies of a building envelope that may cause heat loss and occupant discomfort. Despite its importance, a thermal anomaly is still usually detected by manual analysis of IRT, which strongly depends on the analyzer's experience. In this study, an automatic anomaly detection framework from thermal and visible images was developed. The wall, which is the subject of anomaly detection, is segmented from the visible image by a CNN (Convolutional Neural Network). The temperature threshold of the anomaly area is determined from the multimodal temperature distribution of the target domain. The performance of the anomaly detection was improved by applying the segmentation process (F₁ score 0.497 to 0.808). The framework proposed in this study is expected to be implemented through portable devices and enable instant in-situ thermal anomaly detection.