Envelope 21.195%

Joanne Hastie original
Acrylic on Paper
12" x 16"

This green abstract painting on paper was created as part of my machine learning research. After painting this geometric abstract by hand and not intending it to represent an object. I loaded the image into an artificial intelligence algorithm I use to title the painting (I am using TensorFlow Image Classifier). Of all the potential objects the painting could represent that the algorithm suggested, I selected 'Envelope' as the title, the confidence score is the probability from the algorithm that it could be that object.

Available - contact Joanne to inquire about this painting

Inception v3 results:
envelope (confidence 21.195002)
web site, website, internet site, site (confidence 11.689374%)
umbrella (confidence 2.6394766%)
hatchet (confidence 2.633765%)
pinwheel (confidence 1.8773906%)


I wrote Python code to collage and generate compositions of hand painted geometric shapes and then use an image classifier to rank the compositions based on my preferences.

Over time, I update the classifier as I hone my eye as to what looks good as a painting versus a digital thumbnail. After it is painted, I use an untrained classifier to title the painting. I get titles such as “Bulletproof Vest”, “Traffic Lights” and “Envelope”, because they are flat paintings (minimal depth) – most paintings are classified as “Envelope”. Can you see the objects in the painting?

This work has been featured in NeurIPS Online Gallery in 2018, at ‘Women in Data Science’ show in Zurich in 2018 and in RUSI Journal 2019.

To learn more about this project click below.


Pin It on Pinterest