Deep Learning Neural Networks in Building Superstructure Classification

The conditions of building superstructures require a periodical examination to ensure their reliability and structural health. The current practice of building inspection is carried out using manual visual observation and it leads to subjective decisions across multiple observers. To reduce the confusion of superstructure observation, a deep learning approach is proposed to guide the superstructures learning such as aprons, drains, ceilings and other critical structural components. In this paper, three deep learning approaches are studied for superstructures classification, i.e. VGG16, Vision Transformer (ViT) and Token-to-Token Vision Transformer (T2T-ViT). VGG16 is a deep convolutional neural network having larger trained parameters network. It leads to better accuracy in the task of classification, but the scalability issue is a bottleneck with limited computing power. ViT has a mechanism of self-attention heads by computing relationships among pixels in smaller patches of an image in parallel process at a drastically reduced cost. Due to its scalability, it can be trained on larger image dataset, and its performance can be further improved by increasing the model size and the number of self-attention heads. In the current advancement of ViT, T2T-ViT can obtain a better performance in superstructures classification as compared to the vanilla ViT. However, the current ViT networks fail to perform over conventional CNN models when the training data is limited.