Building facade evaluation using instance segmentation on thermal images

To advance building energy efficiency assessments, Deep Learning (DL) technologies such as image segmentation can be leveraged to expedite evaluations. Previous research has applied instance segmentation to compute U-values of doors, walls, windows and facades, or has applied semantic segmentation to identify thermal anomalies. This work focused on training a DL instance segmentation model on a rudimentary dataset to discern and categorize different facade elements, including windows, doors, walls, floors, and thermal anomalies. The fine-tuned Mask RCNN ResNet-50+FPN model yielded the following segmentation Average Precision (AP) values: 45.6 (mean AP), 65.2 (window glazing), 53.5 (window), 34.9 (wall) and 17.2 (thermal bridge). The bounding box AP for thermal bridges, 28.6, highlights the model’s ability to detect thermal bridges. The results presented in this paper indicate the potential for applying DL instance segmentation to thermal images to identify thermal bridges and estimate U-values using the Infrared Thermovision Technique (ITT). The dataset can be found in https://github.com/gerdac/thermal-images-instance.