Deep learning for detecting building façade elements from images considering prior knowledge

• Main Authors: Zhang, Gaowei, Pan, Yue, Zhang, Limao
• Other Authors: School of Civil and Environmental Engineering
• Format: Article
• Language: English
• Published: 2022
• Subjects: Engineering::Civil engineering; Facade Elements Detection; Deep Learning
• Online Access: https://hdl.handle.net/10356/160749
• Institution: Nanyang Technological University

Building façades elements detection plays a key point role in façade defects detection and street scene reconstruction tasks for sustainable city development. Although the artificial intelligence technology has made a breakthrough in image segmentation, it is nontrivial to directly apply standard deep learning approaches for building façade element detection. The main reason is that the existing semantic segmentation networks have a bad performance in predicting highly regularized shapes. This research develops a hieratical deep learning framework with a symmetric loss function to automatically detect building façade elements from images. The new framework contains two types of attention modules, namely, the spatial attention module, and the channel attention module. A new loss function is designed to integrate prior engineering knowledge, which can be used to force the detection of façade elements (e.g., windows, doors, concrete walls, and sunshades) to have a highly proportionate shape. The effectiveness of the developed approach is demonstrated in two public datasets. Experimental results indicate that the developed deep learning framework with a new loss function outperforms state-of-the-art models significantly, where the achieved a Mean Intersection over Union (IoU) on the Ecole Centrale Paris(ECP) dataset (81.9%) brings an improvement of 11.3% over Fully Convolutional Network (FCN) and 4.0% over Deepfaçade, respectively, and the achieved a Mean IoU on the ArtDeco dataset (85.6%) yields an improvement of 11.4% over FCN and 5.8% over Deepfaçade, respectively. Moreover, the developed approach is more practical and effective to detect regularized façade elements, where the detection of the wall components has an IoU of 93.6% on the ECP dataset and 88.6% on the ArtDeco dataset, respectively. Overall, the contribution from the technical aspect is to develop a hieratical deep learning framework consisting of attention modules together with the newly designed loss function and the prior engineering knowledge. The contribution from the practical aspect is to realize the automatic and accurate detection for various building façade elements in complex environments, which can be potentially helpful for the infrastructure monitoring and maintenance operation.