Application serves for automatic image annotation, mainly for microstock websites like Shutterstock, Fotolia,... It is using several approaches, which can be combined and set-up in config.py:
- Collecting keyword from IPTC of image and parsing keywords from the caption in IPTC metadata of the image
- Detecting objects on the image using Object detection module inception_resnet_v2 trained on OpenImagesV4 dataset
- Classification by BiT-M model trained on ImageNet-21k of image itself and objects detected in previous step
- Completing keywords using MUFIN annotation framework with using keywords from previous steps as seed keywords to improve results
- Clone or download this repository from Github
- Install Python and pip, tested with Python 3.6 and 3.7 other version might cause problems
- Activating virtual environment is recommended, prerequisites might have up to 6 GB, tested using venv in Pycharm IDE.
- Run
setup.py, this will installrequirements.txtas well as download NN models and wordnet stopwords. - Mainly on Windows there might be problem with Tensorflow installation, if this occurs make sure to follow this guide.
- Rename the cloned repository if it is necessary due to "Destination Path Too Long" error.
- Run
annotate.pyfrom terminal or using IDE with images as parameters for example:
python annotate.py /path/*.jpg
python annotate.py /path/01.jpg /path/02.jpg - Run
annotate.pywithout any parameter and select images in the pop-up window - Results of annotation are automatically saved to IPTC of annotated images.
- Feel free to experiment with application parameters in
configuration.pyfile