Bayesian Optimization of Convolutional Neural Networks for Categorizing the Height and Nonregularities of Low to Mid-Rise Buildings in Google Street View

EasyChair Preprint 9289

Abstract

There is a constant endeavor to search for more efficient methods in disaster risk management. This is especially significant for an earthquake-prone country such as the Philippines wherein numerous frameworks have been integrated into common practice and continue to be developed until now. However, physically inspecting these buildings may entail unnecessary effort, thus opening the possibility of the automation of building categorization through virtual images of buildings and the use of convolutional neural networks. With that potential in mind, this research aimed to find an optimal configuration of a convolutional neural network that can classify Google Street View images of buildings in the Greater Metro Manila Area according to their height and detect out-of-plane setbacks, soft stories, split levels, and short columns. The hyperparameters for an optimized ResNet50 network were obtained using Bayesian optimization, and its performance was compared to a base network with training hyperparameters obtained from a past research. A total of 2100 images were obtained, although some of the binary classifications were severely imbalanced. Classification results showed that optimized networks performed better for 3 out of the 5 classifications, but not for the soft story and short columns. The main source of error was associated with the lack of statistical analysis since only the means of evaluation metrics were compared, leaving the possibility of the obtained trials being at tail ends of the multivariate normal distribution. Thus, future studies should apply techniques to mend the class imbalances and perform statistical analysis for a more grounded conclusion on the comparison between the two kinds of networks. Despite these shortcomings, there is still promising potential for Bayesian optimization in creating more efficient automated building categorization, and it remains to be a systematic process in obtaining hyperparameters for any classification task.

Keyphrases: Bayesian optimization, Convolutional Neural Network, building classification, deep learning, height, nonregularities, seismic risk

@booklet{EasyChair:9289,
  author    = {Paolo Giancarlo Guzman and Diocel Harold Aquino and Ammiel Mac Barros and Karlo Daniel Colegio},
  title     = {Bayesian Optimization of Convolutional Neural Networks for Categorizing the Height and Nonregularities of Low to Mid-Rise Buildings in Google Street View},
  howpublished = {EasyChair Preprint 9289},
  year      = {EasyChair, 2022}}