paintings_clustering

Telecom Paris IMA project of clustering unlabelled paintings images.

Structure

Images

Paintings are situated at ./ArtemisArt/

Models

Models are loaded at ./models/

Outputs

Outputs for examples and showcases are sent to ./outputs/

Methods

1st method : Archetypes method

Files used for this method are:

• preprocessing.py (preprocesses the image into a 18001800 image mirrored, then produces two images : downscaled image 224224, most uniform crop of size 224*224 grayish)
• archetype_visualisation.py (implements two features : one to visualize archetypes as images to see the proximity to the paintnigs and the other to project archetypes and paintings on a 2D plane using UMAP to have another way to verify similarities)
• archetype_clustering.py (implements clustering on the archetypes to then classify paintings)
• archetype_style_analysis.py (generates archetypes using the pipline developped in the paper Unsupervised Learning of Artistic Styles with Archetypal Style Analysis, it also has some tests at the end of the file with name = 'main' to try the methods)
• main_archetype.ipynb (notebook that uses every other python file to test the full pipeline)

2nd method : Graph method

Files used for this method are:

• invalid_size.py (deletes images that do not have a max dimension of 1800 pixels)
• preprocessing.py (preprocesses the image into a 18001800 image mirrored, then produces two images : downscaled image 224224, most uniform crop of size 224*224 grayish)
• features_extractor.py (retrieves the features of an image: the 5 convolutional activation layers of the downscaled image passed into resnet, and the 5 convolutional activation layers of the crop passed into resnet)
• features_comparator.py (implements a weighted cossine dissimilarity distance measure between to sets of features)
• knn.py (creates a non-complete graph where each image is a node and has only k neighbors which are expected to be its closest neighbors in the theorical complete graph of distances)
• hdbscan_clustering.py (uses hdbscan to compute classes based on the non-complete graph computed in knn)

Each function has some tests in if name=='main' section. To try the full method, use python hdbscan_clustering.py and wait. The classes are outputed in ./outputs/hdbscan_classes/classx/