Associative Trails
← Writing

Associative Trails / writing

OPTIMISING IMAGE CLASSIFICATION - Implementation of Convolutional Neural Network Algorithms to Distinguish Between Plans and Sections within the Architectural, Engineering and Construction (AEC) Industry (2019)

Notes on a 2019 paper by Alessandra Fabbri, M. Hank Haeusler, and Yannis Zavoleas

An attempt to build a convolutional neural network to classify 1000 images into 3 categories. Accuracy peaked at 82%, no hyperparameter tuning was performed.

Sheet 03 · Practice
Published30 Apr 2022
Reading1 min
01
The paperFabbri, Haeusler & Zavoleas · 2019
ResearchGate PDF 1.5 MB
  • Build a CNN to classify drawings into sections, plans and elevations
  • 1,000 images used - 800 for training, 200 for model validation
  • Objective to achieve accuracy of 90% or more.
  • Gives lots of detail about the different iterations of models. First peaked at 52% accuracy over 128 epochs. Second peaked at 78% after increasing the number of convolutional layers. Third lowered the learning rate and peaked at 82%.
  • They recommend further adjustment of hyperparameters to improve the results, with more trainman images. Also consider looking at alternative CNN architectures - highway networks (Srivastava et al. 2015), residual networks (He et al. 2016), dense networks (Huang et al. 2017)
  • Deploy as executable or web application.

We help professional services firms work out where AI genuinely helps, what has to be structured first, and what to leave alone.

Get in touch