Lab-scale termite damage synthesis using least squares generative adversarial networks

IRG/WP 20-20674

D J V Lopes

This manuscript investigated the feasibility of least squares generative adversarial networks (LSGAN) to generate synthetic images of lab-scale termite damage based on AWPA E1 standard, to push machine-learning forward into wood science field and to ameliorate the lack of a termite damage dataset. We leveraged LSGAN to learn the distribution of 203 uniquely termite damaged samples from previous experiments. Our novel approach generated 256 x 256 pixels images with realistic and meaningful details. However, checkerboard artifacts were generated. The Fréchet Inception Distance (FID) was calculated and achieved averaged value below 25. This approach could be used to enhance previously done image classification model on termite damage severity. Future research will explore using GAN generated images to augment small datasets of real images and quantify any increases of image classification robustness.


Keywords: machine-learning, generative models, termites, lab-scale damage, AWPA E1

Conference: 20-06-10/11 IRG51 Webinar


Download document (371 kb)
free for the members of IRG. Available if purchased.

Purchase this document