Workflow for Window Composition Detection to Aid Energy-Efficient Renovation in Low-Income Housing in Korea

Abstract

Enhancing the efficiency of windows is important for improving the energy efficiency of buildings. The Korean government has performed numerous building renovation projects to reduce greenhouse gas emissions and mitigate energy poverty. To reduce the costs and manpower requirements of conventional field surveys, this study presents a deep-learning model to examine the insulation performance of windows using photographs taken in low-income housing. A smartphone application using crowdsourcing was developed for data collection. The insulation performance of windows was determined based on U-value, derived considering the frame-material type, number of panes, and area of windows. An image-labeling tool was designed to identify and annotate window components within photographs. Furthermore, software utilizing open-source computer vision was developed to estimate the window area. After training on a dataset with ResNet and EfficientNet, an accuracy of approximately 80% was achieved. Thus, this study introduces a novel workflow to evaluate the insulation performance of windows, which can support the energy-efficient renovation of low-income housing.