Detecting Text Baselines in Historical Documents With Baseline Primitives

Previous deep learning based approaches to text baseline detection in historical documents usually take it as a semantic segmentation task. These methods adopt a fully convolutional neural network to predict baseline pixels first and then group them into lines by heuristic post-processing steps, which tends to suffer from a wrongly merged or wrongly split problem owing to limited context information provided by pixels. To address these issues, we introduce the concept of a baseline primitive, which is defined as a virtual bounding box centered at each baseline pixel. After baseline primitive detection, a relation network is used to predict a link relationship for each pair of primitives. Consequently, text baselines are generated by detecting baseline primitives and grouping them with the corresponding link relationships. Owing to the design of baseline primitives, wider context information can be leveraged to improve link prediction accuracy. Therefore, our approach can effectively detect text baselines with small inter-line or large inter-word spacing. Quantitative experimental results demonstrate the effectiveness of the proposed baseline primitive design. Our approach achieves state-of-the-art performance on two public benchmarks, namely cBAD 2017 and cBAD 2019.