PT Unknown
AU Böschen, F
TI Analysis and Modular Approach for Text Extraction from Scientific Figures on Limited Data
PY 2021
PU Christian-Albrechts-Universität zu Kiel
WP https://macau.uni-kiel.de/receive/macau_mods_00002238
LA en
DE Text Extraction; Infographic; Scientific Figure
AB Scientific figures are widely used as compact, comprehensible representations of important information. The re-usability of these figures is however limited, as one can rarely search directly for them, since they are mostly indexing by their surrounding text (e. g., publication or website) which often does not contain the full-message of the figure. In this thesis, the focus is on making the content of scientific figures accessible by extracting the text from these figures. A modular pipeline for unsupervised text extraction from scientific figures, based on a thorough analysis of the literature, was built to address the problem. This modular pipeline was used to build several unsupervised approaches, to evaluate different methods from the literature and new methods and method combinations. Some supervised approaches were built as well for comparison. One challenge, while evaluating the approaches, was the lack of annotated data, which especially needed to be considered when building the supervised approach. Three existing datasets were used for evaluation as well as two datasets of 241 scientific figures which were manually created and annotated. Additionally, two existing datasets for text extraction from other types of images were used for pretraining the supervised approach. Several experiments showed the superiority of the unsupervised pipeline over common Optical Character Recognition engines and identified the best unsupervised approach. This unsupervised approach was compared with the best supervised approach, which, despite of the limited amount of training data available, clearly outperformed the unsupervised approach.
PI Kiel
ER