Copies and Originals Project

May 25, 2016

Undergraduate Bridget Willke in the Department of Statistical Science worked on her senior Statistics honors thesis with Dr. Ingrid Daubechies on how analysis of brushstrokes might be used to distinguish copies from originals.

In addition to being a Statistics major, Bridget also has a minor in Art History. She was especially excited to develop a senior thesis project that would combine both of these interests in addressing how to distinguish original paintings from copies. Bridget’s project is based on an algorithm that uses techniques from machine learning, mathematical modeling and image processing to detect the difference in digitized paintings.

Envolope opening

“Dr. Daubechies and Dr. Dunson’s guidance helped me to not only gain a better understanding of wavelets and statistical modeling, but also challenged me to pursue my curiosity and look for solutions on my own,” Bridget says. “I am extremely grateful for the opportunity to have worked with such wonderful advisors, and I look forward to expanding on the work over the summer with Dr. Daubechies.”

Data extracted from each scanned painting with Dual-Tree Wavelet transforms produced features for different patches of the painting. These features and the underlying structure of the image were captured using a Hidden Markov Tree model. Support Vector Machines with different kernel functions, following methods from a previous experiment by a Princeton team (Polatkan et al., 2009), were used to perform classification.

“If you have a style you’re comfortable with, your brush strokes will look different and actually leave a trace when you start looking at these building blocks of the image than if you tried to copy or imitate the style,” Dr. Daubechies explained.

The final test was to apply Bridget’s algorithms to an original painting, and to a copy by the same artist. iiD hosted an event where everyone was invited to predict which painting was which, along with enjoying some cake and ice cream!

Nobody knew the answer until our sponsor Aziz Ahmadieh opened the envelope, and to Bridget’s delight, she discovered her algorithm had predicted correctly.

Predictions among the human participants at the event favored the copy over the original, however. This bodes well for the future Bridget’s successful work—congratulations, Bridget!