A deep learning method for building façade parsing utilizing improved SOLOv2 instance segmentation

Abstract

Energy consumption simulation and renovation of existing buildings require accurate acquisition of building façade features which mostly relies on time-consuming manual calculations based on architectural drawings. In this article, we proposed an automated deep learning-based approach based on the SE module and BiFPN to achieve precise and efficient façade feature extraction. The approach eliminated the image distortion of building façades and then enabled accurate segmentation of windows and accessory structures even under the situation of occlusion and reflection. The improved SOLOv2 algorithm resulted in a high mean average precision of 93% for window segmentation, leading to a more precise window-to-wall ratio estimation with a mean absolute error of 2.9% than the experts’ estimation and existing deep learning-based methods. Considering the accurate results of façade parsing, our method can be utilized for city-level building feature extraction, providing theoretical and practical references for urban building energy simulation, urban renewal, and building health examination.