Abstract by Madeline Reinhard

Personal Infomation

Presenter's Name

Madeline Reinhard

Degree Level


Abstract Infomation


Computer Science

Faculty Advisor

Sean Warnick


Objective Art: Analyzing Convolutional Neural Networks’ Ability to Identify Objects from Realistic to Abstract Paintings


The aim of this work is to model the degradation of a convolutional neural network trained on photographs as it identifies objects in increasingly abstract paintings. A network trained to identify objects in various painting styles will provide art historians a tool to navigate large amounts of visual data. Due to a lack of labeled data, it is difficult to train a network that can identify painted objects with sufficient accuracy.

From a dataset of 9000 paintings, I will derive three levels of abstractness and take a network pre-trained on photos and evaluate its effectiveness at identifying objects in each group. I will retrain the final layer of the network on each of the groups and test it against photographs and the other two groups of paintings. I will repeat the process for each group to see if there is a significant difference in accuracy. This will hopefully provide an estimate of how much labeled data is required to train a network to identify objects in art.