@phdthesis{Bartz2022,
  author    = {Bartz, Christian},
  title     = {Reducing the annotation burden: deep learning for optical character recognition using less manual annotations},
  doi       = {10.25932/publishup-55540},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-555407},
  school      = {Universit{\"a}t Potsdam},
  pages     = {xxiv, 183},
  year      = {2022},
  abstract  = {Text is a ubiquitous entity in our world and daily life. We encounter it nearly everywhere in shops, on the street, or in our flats. Nowadays, more and more text is contained in digital images. These images are either taken using cameras, e.g., smartphone cameras, or taken using scanning devices such as document scanners. The sheer amount of available data, e.g., millions of images taken by Google Streetview, prohibits manual analysis and metadata extraction. Although much progress was made in the area of optical character recognition (OCR) for printed text in documents, broad areas of OCR are still not fully explored and hold many research challenges. With the mainstream usage of machine learning and especially deep learning, one of the most pressing problems is the availability and acquisition of annotated ground truth for the training of machine learning models because obtaining annotated training data using manual annotation mechanisms is time-consuming and costly. In this thesis, we address of how we can reduce the costs of acquiring ground truth annotations for the application of state-of-the-art machine learning methods to optical character recognition pipelines. To this end, we investigate how we can reduce the annotation cost by using only a fraction of the typically required ground truth annotations, e.g., for scene text recognition systems. We also investigate how we can use synthetic data to reduce the need of manual annotation work, e.g., in the area of document analysis for archival material. In the area of scene text recognition, we have developed a novel end-to-end scene text recognition system that can be trained using inexact supervision and shows competitive/state-of-the-art performance on standard benchmark datasets for scene text recognition. Our method consists of two independent neural networks, combined using spatial transformer networks. Both networks learn together to perform text localization and text recognition at the same time while only using annotations for the recognition task. We apply our model to end-to-end scene text recognition (meaning localization and recognition of words) and pure scene text recognition without any changes in the network architecture. In the second part of this thesis, we introduce novel approaches for using and generating synthetic data to analyze handwriting in archival data. First, we propose a novel preprocessing method to determine whether a given document page contains any handwriting. We propose a novel data synthesis strategy to train a classification model and show that our data synthesis strategy is viable by evaluating the trained model on real images from an archive. Second, we introduce the new analysis task of handwriting classification. Handwriting classification entails classifying a given handwritten word image into classes such as date, word, or number. Such an analysis step allows us to select the best fitting recognition model for subsequent text recognition; it also allows us to reason about the semantic content of a given document page without the need for fine-grained text recognition and further analysis steps, such as Named Entity Recognition. We show that our proposed approaches work well when trained on synthetic data. Further, we propose a flexible metric learning approach to allow zero-shot classification of classes unseen during the network's training. Last, we propose a novel data synthesis algorithm to train off-the-shelf pixel-wise semantic segmentation networks for documents. Our data synthesis pipeline is based on the famous Style-GAN architecture and can synthesize realistic document images with their corresponding segmentation annotation without the need for any annotated data!},
  language  = {en}
}
@book{BartzKrestel2021,
  author    = {Bartz, Christian and Krestel, Ralf},
  title     = {Deep learning for computer vision in the art domain},
  number    = {139},
  publisher = {Universit{\"a}tsverlag Potsdam},
  address   = {Potsdam},
  isbn      = {978-3-86956-514-9},
  issn      = {1613-5652},
  doi       = {10.25932/publishup-51290},
  url       = {http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-512906},
  publisher      = {Universit{\"a}t Potsdam},
  pages     = {vii, 79},
  year      = {2021},
  abstract  = {In recent years, computer vision algorithms based on machine learning have seen rapid development. In the past, research mostly focused on solving computer vision problems such as image classification or object detection on images displaying natural scenes. Nowadays other fields such as the field of cultural heritage, where an abundance of data is available, also get into the focus of research. In the line of current research endeavours, we collaborated with the Getty Research Institute which provided us with a challenging dataset, containing images of paintings and drawings. In this technical report, we present the results of the seminar "Deep Learning for Computer Vision". In this seminar, students of the Hasso Plattner Institute evaluated state-of-the-art approaches for image classification, object detection and image recognition on the dataset of the Getty Research Institute. The main challenge when applying modern computer vision methods to the available data is the availability of annotated training data, as the dataset provided by the Getty Research Institute does not contain a sufficient amount of annotated samples for the training of deep neural networks. However, throughout the report we show that it is possible to achieve satisfying to very good results, when using further publicly available datasets, such as the WikiArt dataset, for the training of machine learning models.},
  language  = {en}
}