gfolego / vangogh

From Impressionism to Expressionism: Automatically Identifying Van Gogh's Paintings

Automatically Identifying Van Gogh's Paintings

This is the source code used in the paper "From Impressionism to Expressionism: Automatically Identifying Van Gogh's Paintings", which has been published on the 23rd IEEE International Conference on Image Processing (ICIP 2016).

The paper is available at IEEE Xplore: https://dx.doi.org/10.1109/icip.2016.7532335

The dataset is available at figshare: https://dx.doi.org/10.6084/m9.figshare.3370627

Corresponding author: Anderson Rocha ([e-mail])

If you find this work useful in your research, please cite the paper! :-)

@InProceedings{folego2016vangogh,
    author = {Guilherme Folego and Otavio Gomes and Anderson Rocha},
    booktitle = {2016 IEEE International Conference on Image Processing (ICIP)},
    title = {From Impressionism to Expressionism: Automatically Identifying Van Gogh's Paintings},
    year = {2016},
    month = {Sept},
    pages = {141--145},
    keywords = {Art;Feature extraction;Painting;Support vector machines;Testing;Training;Visualization;CNN-based authorship attribution;Painter attribution;Data-driven painting characterization},
    doi = {10.1109/icip.2016.7532335}
}

---

Quick Guide

This guide has four sections:

1. Creating the dataset - Create your own dataset, or (even better!) expand on VGDB-2016.
2. Using our method - Use our method, given a dataset.
3. Predicting debated paintings - Predict debated paintings with our method.
4. Calculating scores - Transform distances into probabilities.

General note: all the scripts presented here have a --help argument, which describes the script and possible parameters.

Creating the dataset

Requirements (for this section)

• ImageMagick
• Python, and the following packages:
  • hurry.filesize
  • numpy
  • progressbar2
  • wikitools
• R, and the following packages:
  • argparse
  • data.table
  • dplyr

Create a directory for resources.

mkdir -pv res/{db,img/{orig,resz}}

Define the URL to be crawled. This is just an example. In our work, we crawled more than 200 different URLs.

url='Category:Still_life_paintings_of_flowers_by_Vincent_van_Gogh,_Auvers_1890'

Crawl URL and collect metadata.

python src/crawler/crawl2csv.py --url "$url" --csv res/db/"$url"

Parse and clean up collected metadata. We set different values here just as a working example. Also, at this point, it is possible to provide multiple files at once, even with duplicated entries (as shown).

Rscript src/crawler/tidy_dataset.R --density 95 --ratio 0.15 --output res/db/db.csv res/db/"$url" res/db/"$url"

Dataset done, and the CSV file is at res/db/db.csv. Now, you may choose to continue with your newly created dataset, or with the original vgdb_2016.csv.

Download images.

python src/crawler/download_images_from_csv.py --csv res/db/db.csv --directory res/img/orig/

Note: images with less than 75% of JPEG quality were manually removed (both images and respective entries in the CSV file). It is possible to check the quality with ImageMagick.

identify -format "%f:%Q\n" res/img/orig/* | grep -v ^$ | sort -k2nr -k1 -t:

Resize images to the standard density.

python src/crawler/resize_images.py --csv res/db/db.csv --original res/img/orig/ --resized res/img/resz/

Using our method

Requirements (for all the following sections)

• Caffe
• Parallel
• Python, and the following packages:
  • scikit-image
  • scikit-learn
• Unzip

From now on, we will assume that the vgdb_2016.zip dataset file has already been downloaded.

Unzip the dataset.

unzip vgdb_2016.zip

Create a directory for resources.

mkdir -pv vgdb_2016/{train,test}/{patch,feats}

Extract patches from each image.

find vgdb_2016/train/{,n}vg -type f | parallel python src/analysis/patch_extraction.py --image {} --dir vgdb_2016/train/patch/
find vgdb_2016/test/{,n}vg -type f | parallel python src/analysis/patch_extraction.py --image {} --dir vgdb_2016/test/patch/

Extract features from each patch. In our work, we used the VGG 19-layer model, which is available at http://www.robots.ox.ac.uk/~vgg/research/very_deep/.

ls vgdb_2016/train/patch/ > vgdb_2016/train/patch_list.txt
ls vgdb_2016/test/patch/ > vgdb_2016/test/patch_list.txt

python src/analysis/caffe_extract_features.py --proto path/to/VGG_ILSVRC_19_layers_deploy.prototxt --model path/to/VGG_ILSVRC_19_layers.caffemodel --list vgdb_2016/train/patch_list.txt --input vgdb_2016/train/patch/ --output vgdb_2016/train/feats/
python src/analysis/caffe_extract_features.py --proto path/to/VGG_ILSVRC_19_layers_deploy.prototxt --model path/to/VGG_ILSVRC_19_layers.caffemodel --list vgdb_2016/test/patch_list.txt --input vgdb_2016/test/patch/ --output vgdb_2016/test/feats/

Create a directory for the classification model.

mkdir -pv vgdb_2016/clf

Generate classification model.

python src/analysis/generate_model.py --dir vgdb_2016/train/feats/ --model vgdb_2016/clf/model.pkl

Classify paintings in the test set using the Far method.

python src/analysis/classify.py --dir vgdb_2016/test/feats/ --model vgdb_2016/clf/model.pkl --aggregation far --gtruth

Done!

Predicting debated paintings

Create a directory for resources.

mkdir -pv vgdb_2016/check/{patch,feats}

Extract patches from each image.

find vgdb_2016/check/[0-9]*.png -type f | parallel python src/analysis/patch_extraction.py --image {} --dir vgdb_2016/check/patch/

Extract features from each patch.

ls vgdb_2016/check/patch/ > vgdb_2016/check/patch_list.txt
python src/analysis/caffe_extract_features.py --proto path/to/VGG_ILSVRC_19_layers_deploy.prototxt --model path/to/VGG_ILSVRC_19_layers.caffemodel --list vgdb_2016/check/patch_list.txt --input vgdb_2016/check/patch/ --output vgdb_2016/check/feats/

Classify paintings using the Far method.

python src/analysis/classify.py --dir vgdb_2016/check/feats/ --model vgdb_2016/clf/model.pkl --aggregation far

In the output, class 1 means van Gogh, and class 0 means non-van Gogh.

Calculating scores

Generate scores model.

python src/analysis/generate_score_model.py --dir vgdb_2016/train/feats/ --model vgdb_2016/clf/model.pkl --score vgdb_2016/clf/score.pkl

Calculate score probabilities. Targets are the filenames without extension, separated by comma.

targets='9414428,9420113'
echo $targets | sed s/,/\\n/ > vgdb_2016/check/target_list.txt
python src/analysis/get_scores.py --dir vgdb_2016/check/feats/ --model vgdb_2016/clf/model.pkl --score vgdb_2016/clf/score.pkl --targets vgdb_2016/check/target_list.txt

In the output, the first column represents non-van Gogh, and the second column represents van Gogh.