{ Learning Mid-Level Image Features to Guide Evolutionary Art Generation }


Figure 1. Curated set of evolutionary artworks generated by Professor Andrews' codes based on genetic algorithms.

The computational evolutionary art generating process is similar to the biological evolution in which cross-over and mutation of operators happen in a random fashion, which then dictates the content of the generated image. Users can guide the generation by selecting certain images to breed new generations.


The goal of this research project is to build a neural network that can recognize the level of certain mid-level features in order to spatialize these images in a meaningful coordinate system (with axis representing the level of curviness, complexity, or color smoothness in an image). The ultimate goal is to provide evolutionary artists with more control over the selection process.

Besides building the neural network, we also built a web application survey to elicit parameters and descriptive words from people and to gain insights into how to best represent abstract images like these evolutionary artworks. 

For more details, see posters.