DOI: https://doi.org/10.15368/theses.2018.69
Available at: https://digitalcommons.calpoly.edu/theses/1934
Date of Award
6-2018
Degree Name
MS in Computer Science
Department/Program
Computer Science
Advisor
Maria Pantoja
Abstract
There are currently different efforts to use Supervised Neural Networks (NN) to automatically label damages on images of above ground infrastructure (buildings made of concrete) taken after an earthquake. The goal of the supervised NN is to classify raw input data according to the patterns learned from an input training set. This input training data set is usually supplied by experts in the field, and in the case of this project, structural engineers carefully and mostly manually label these images for different types of damage. The level of expertise of the professionals labeling the training set varies widely, and some data sets contain pictures that different people have labeled in different ways when in reality the label should have been the same. Therefore, we need to get several experts to evaluate the same data set; the bigger the ground truth/training set the more accurate the NN classifier will be. To evaluate these variations among experts, which can be considered equal to the task of evaluating the quality of the expert, using probabilistic theory we first need to implement a tool able to compare different images classified by different experts and apply a certainty level to the experts tagged labels. This master's thesis implements this comparative tool. We also decided to implement the comparative tool using parallel programming paradigms since we foresee that it will be used to train multiple young engineering students/professionals or even novice citizen volunteers (“trainees”) during after-earthquake meetings and workshops. The implementation of this software tool involves selecting around 200 photographs tagged by an expert with proven accuracy (“ground truth”) and comparing them to files tagged by the trainees. The trainees are then provided with instantaneous feedback on the accuracy of their damage assessment. The aforementioned problem of evaluating trainee results against the expert is not as simple as comparing and finding differences between two sets of image files. We anticipate challenges in that each trainee will select a slightly different sized area for the same occurrence of damage, and some damage-structure pairs are more difficult to recognize and tag. Results show that we can compare 500 files in 1.5 seconds which is an improvement of 2x faster compared to sequential implementation.