I chose to classify impressionist painters by their paintings. My model includes work from impressionist painters Claude Monet, Vincent Van Gogh, Camille Pissarro, and Mary Cassatt. I trained it to identify who, of the four included, painted the painting that is uploaded to be identified. I wanted to classify impressionist painters because I think it is interesting how painters’ work can be categorized into a genre while the painter still maintains their own unique style within that genre to make their work identifiable. There are also some cases where painters’ work is very similar to each other. I wanted to see how the Teachable Machine Program would handle deciphering painters who paint very similar subjects, versus painters whose typical subjects differ.
In building the classification model, I chose images from the Wikipedia Commons page published for each individual artist. When I was picking images of paintings to include, I tried to choose a wide variety of work ranging from portraits of people to landscapes in nature. I wanted to give the program examples of as many of the different focal points that artists paint as possible. I also intentionally chose to classify some impressionist artists that produce work that is similar enough to be confused for each other because I wanted to see how the program would respond.
My model was successful in identifying paintings that were very clear in who painted them based on subject and style. When I tested an image of a painting that looked similar to work from multiple artists, the model was not always confident or even correct. In some cases, the model would report partial probability for more than one artist, and sometimes the artist that it thought most likely painted the painting was not correct at all. I think this is partially due to the fact that there are a lot of similarities in subjects and painting styles of the artists I chose. It would be difficult for the model to determine who painted it with absolute certainty given the provided information having such overlapping themes. To make the model more successful, I added an additional image for each artist. The more examples provided, the more information the model has to rely on, and thus the more accuracy it has when identifying unknown art works. To improve accuracy, I also tried to make sure the additional image was a different kind of example. This would hopefully broaden the model’s pool of information it can use when attempting to identify paintings it is not familiar with.
This model could be used in a heritage experience that categorizes paintings based on the artist and their heritage. It could be used create an extensive collection of art from the histories of various cultures. Additionally, the model could help classify artists based on their work if it is not known what genre they fit into. For example, if we input their paintings into the model and the model reports that they are likely created by a known impressionist painter then we can identify that artist as an impressionist painter. We could also use the model to add a new artist to the exhibit and categorize them with others from the same heritage if that information is unknown.