Development of a semantic segmentation network for plane layout design of super high-rise building structures

Development of intelligent structural design (ISD) of buildings has gained increasing attention. The need of ISD for complex super high-rise buildings with inherent substantial challenges is particularly more pressing. As the fundamental step of ISD, a properly-designed semantic segmentation network is essential for extracting the pixel-level sparse information contained in the dramatically variable structural plane layouts along height. As the result of the variations, the segmentation networks developed to date fail to achieve this goal both accurately and efficiently. The well-structured DeepLabv3+ is chosen as the baseline network on which some significant modifications are made to develop Tall-DeepLabv3+, i.e. the replacement of the encoding backbone with a multi-scale convolutional attention network and the integration of three external attention modules, three skip connections and a unified loss function. The accuracy, stability and generalization ability of Tall-DeepLabv3+ is systematically demonstrated through the pre-training, transfer training, ablation experiments and comparative validation analysis. Utilizing a parametrically-generated dataset for frame core-tube building structures, the network achieved a peak mean intersection over union of 93.52 % with a minimal deviation of 0.002. Comparative validation analysis results demonstrated the superior overall performance and segmentation type-level accuracy of Tall-DeepLabv3+ in processing plane layouts characterized by prominent sparsity features.