Date of Graduation

5-2018

Document Type

Thesis

Degree Name

Bachelor of Science

Degree Level

Undergraduate

Department

Computer Science and Computer Engineering

Advisor

Gauch, John

Reader

Li, Qinghua

Second Reader

Gashler, Michael

Abstract

In this paper, we compare the results of ResNet image classification with the results of Google Image search. We created a collection of 1,000 images by performing ten Google Image searches with a variety of search terms. We classified each of these images using ResNet and inspected the results. The ResNet classifier predicted the category that matched the search term of the image 77.5% of the time. In our best case, with the search term “forklift”, the classifier categorized 92 of the 100 images as forklifts. In the worst case, for the category “hammer”, the classifier matched the search term 61 times out of 100. We also leveraged the prediction confidence levels of the ResNet classifier to determine the relative similarity of an image within a set of images. In typical usage of an image classifier, only the most confident prediction is utilized. By using a larger piece of the output vector of the ResNet classifier, we were able to calculate distances between images in feature space. We created visualizations of the distance between images in sets of 100 images.

Keywords

computer vision, convolutional neural network, neural network, AI, image classification

Share

COinS