Image Collection Visualizer

Visualizing AMNH's Photographic Collection with machine learning and Library metadata.

Requirements

• Python (This is developed using 3.6, so 3.6+ is recommended and may not work with 2.7+)
• SciPy for math functions (probably already installed)
• Keras for image feature extraction
• Scikit-learn for feature reduction (e.g. PCA)
• Multicore-TSNE for converting features to 2 dimensions via TSNE
• RasterFairy for transforming 2D points to grid
• Pillow for image tile generation
• Node.js if you'd like to run the interface locally

Workflow

First, given a directory of images, we will extract 4096 features using Keras and the VGG16 model with weights pre-trained on ImageNet, then reduce those to 256 features using PCA, then save those features to a compressed file:

python images_to_features.py \
-in "images/photographic_thumbnails/*.jpg" \
-pca 256 \
-out "output/photographic_features.p.bz2"

Then we will reduce those features even further to just two dimensions using TSNE and output the result to a csv file. You can speed this up by indicated the number of parallel jobs to run, e.g. -jobs 4

python features_to_tsne.py \
-in "output/photographic_features.p.bz2" \
-jobs 4 \
-out "data/photographic_tsne.csv"

Then we will convert those 2D points to a grid assignment using RasterFairy. Note that Rasterfairy only supports Python 2.x as of this writing.

python tsne_to_grid.py \
-in "data/photographic_tsne.csv" \
-out "data/photographic_grid.csv"

We will then generate a giant image matrix from the images and the grid data using a 128x128 tile size and 114x116 target grid size:

python grid_to_image.py \
-in "data/photographic_grid.csv" \
-tile "128x128" \
-grid "114x116" \
-out "output/photographic_matrix.jpg"

Finally, we will convert the giant image to tiles (in .dzi format):

python image_to_tiles.py \
-in "output/photographic_matrix.jpg" \
-tsize 128 \
-out "img/photographic_matrix.dzi"

If you have metadata and subjects in .csv format like this file and this file, you can convert it to .json for it to be used by the interface. Note the id column must match the associated image filename (without extension). The grid data and grid size is also indicated so the metadata would align properly:

python csv_to_json.py \
-in "data/photographic_images.csv" \
-sub "data/photographic_subjects.csv" \
-image "images/photographic_thumbnails/*.jpg" \
-grid "data/photographic_grid.csv" \
-gsize "114x116"
-out "data/photographic_images.json"

You can view the result on a local server by running:

npm install
npm start