To make an artstyle detector, I first had to find enough images for the training. The following are the list of the sources I used or at least tried to use to gather a bunch of images.

# i : input image, o : model path python src/main.py -i tests/monet_impressionism.jpg -o styles/ >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> Predictions for image: tests/monet_impressionism.jpg ============================================================ neo_impressionism : 48.7353 impressionism : 46.3269 post_impressionism : 1.6898 pre_raphaelite_brotherhood : 1.5794 naive_art_primitivism : 0.4833 realism : 0.3755 expressionism : 0.3007 romanticism : 0.2878 surrealism : 0.153 abstract_expressionism : 0.0682 ============================================================

The first approach may not always be the best (it's never). To crawl Google Images I had to use Google API, make a custom Programmable Search engine and also be under a limit of requests. I didn't look at it further but made a test run.

Example of wikimedia categories Impressionist_paintings. The way I crawled this was pretty stupid but honest (bs4s, asyncio classic). Why? I could try an approach using SPARQL queries at Wikidata Query Service. I may implement it in the future.

As the wikimedia images were not enough ?!, I tried another approach. I found Wikiart which seemed to have an adequate amount of images at first glance. Luckily, there was a repository working on this, so I downloaded the datasets by the links provided.

To train the model I used ImageAI. After writing some functionalities to construct the directory structure, training was pretty straightforward.

I was not satisfied with the accuracy it achieved but it was probably because of the dataset similarities. I'll try to use more categories in the future.

Truth is, a painting can have multiple arstyles. I tested some pictures, as welll as some of my own, and the results were okayish, but with not high confidence most of the times.

I used color-thief to extract paletters and dominant colors from the images. I am still not sure what to do with this information but it was cool. I thought about organizing the images to clusters according to their palette range.