Seminar Knowledge Mining

Wikimedia image classification and suggestions for article authors.

Set up instructions

Unix

1. Install these dependencies by using your system's package manager if you don't have them already.

   Depdendency	Apt	Pacman	Homebrew
   Python 3	python3	python
   Cython	cython3	cython
   Pip	python3-pip	python-pip
   Virtualenv	virtualenv	python-virtualenv
   Fortran	gfortran	gcc-fortran
   Blas	libblas-dev	blas
   Lapack	liblapack-dev	lapack
   PNG	libpng-dev	libpng
   JPEG	libjpeg8-dev	libjpeg-turbo
   Freetype	libfreetype6-dev	freetype2
   Cairo	libcairo2-dev		cairo
   FFI	libffi-dev

2. Create a virtual environment inside the repository root by runnning virtualenv . or if you have multiple Python versions virtualenv -p python3 ..

3. Activate your virtual environment using source bin/activate. Make sure that the repository name is in front of your shell promt now.

4. Install dependencies inside your virtual environment

    pip install -r requirements.txt
   

5. Install OpenCV 3.0 with bindings for Python 3 by running

    chmod +x tool/setup-opencv.sh
    tool/setup-opencv.sh
   

6. UTF-8 is required, so you may need to add these lines to your ~/.bash_profile and apply the changes with source ~/.bash_profile.

    export LC_ALL=en_US.UTF-8
    export LANG=en_US.UTF-8
   

Windows

1. Create a virtual environment inside the repository root by runnning virtualenv . or if you have multiple Python versions virtualenv -p C:\Python34\python.exe ..

2. Activate your virtual environment using Scripts\activate. Make sure that the repository name is in front of your shell promt now.

3. Download these dependencies. If in doubt, use the link before the last in each list. Run pip install <path-to-file> on each of those.

   • NumPy
   • SciPy
   • Scikit-learn
   • Scikit-image

4. Install remaining dependencies inside your virtual environment using pip install -r requirements.txt.

Workflows

Data set

1. Download DBpedia dump
2. Extract list of image names
3. Fetch image and meta data of random entries
4. Manually label data
5. Balance amount of image per class

Training

1. Preprocess data set
2. Extract image and text based features
3. Train classifier

Image classification

1. Get user search term
2. Query DBpedia for related images based on description
3. Fetch image and meta data of first results
4. Extract image and text based features
5. Use trained classifier to predict class
6. Store results in DBpedia

Usage

-h can be passed as a parameter to get a comprehensive list of parameters for all the classes listed below.

fetch_[source].py files can be used to download images for a test/training set from [source] to test the classifier.

extraction.py can be used to extract textual and visual features from images.

classifier.py can be used to classify images.

performance.py trains a different classifiers and measures their individual performance.

evaluation.py can be used to measure the performance of individual features.