The goal of this project is to develop automatic image tagging methods. The first phase was creating a large dataset of images belonging to different semantic categories (e.g. product images, logos, indoor recordings and outdoor photos). We will now investigate deep-learning-based methods for determining the characteristics of the different semantic categories. These methods will be trained on our dataset and then used to automatically predict semantic labels (tags) for new images. We will use pre-trained neural networks, which we will retrain and adapt the architecture of in order to perform the necessary domain adaptation.
Project partner: mediamid digital services GmbH