Compression-Based Digital Watermarking of Color Images Using Quaternion-Valued Neural Networks — Presentation Detail — Celebration of Scholars

Additional Information

More information is available at carthage.edu/celebration-scholars/. The following are members of the Research, Scholarship, and Creativity Committee who are eager to listen to ideas and answer questions:

Thomas Carr
Katherin Hilson
Kim Instenes
John Kirk
Sarah Terrill

Compression-Based Digital Watermarking of Color Images Using Quaternion-Valued Neural Networks

Name: Riley Maguire
Major: Mathematics
Hometown: Grayslake, IL
Faculty Sponsor: Diana Thomson
Other Sponsors:
Type of research: SURE
Funding: SURE

Name: Alec DiGirolamo
Major: Computer Science
Hometown: Pleasant Prairie, WI
Faculty Sponsor: Diana Thomson
Other Sponsors:
Type of research: SURE
Funding: SURE

Abstract

We develop a neural network to add a digital watermark to an image. This watermark changes how the color of the image is processed, to color images with a network that uses quaternions. Quaternions are 4-dimensional vectors with three complex bases and a real part. In order to make the network classify data as accurately as possible, we minimize a function that determines the inaccuracy, called the error function. Since quaternions are non-commutative under multiplication, we also determine the most efficient multiplication order for minimizing the error function of the neuron. While many have investigated how to make neural networks more efficient, improving network efficiency by changing the multiplication order of quaternions has not been as thoroughly researched. Here, we present the most efficient multiplication order of quaternions, as well as a network that implements digital watermarks.

Submit date: March 18, 2020, 4:45 p.m.