Ce Zhang

Information in industry, research, and the public sector is widely stored as rendered documents (e.g., PDF files, scans). Hence, to enable downstream tasks, systems are needed that map rendered documents onto a structured hierarchical format. However, existing systems for this task are limited by heuristics and are not end-to-end trainable. In this work, we introduce the Document Structure Generator (DSG), a novel system for document parsing that is fully end-to-end trainable. DSG combines a deep neural network for parsing (i) entities in documents (e.g., figures, text blocks, headers, etc.) and (ii) relations that capture the sequence and nested structure between entities. Unlike existing systems that rely on heuristics, our DSG is trained end-to-end, making it effective and flexible for real-world applications. We further contribute a new, large-scale dataset called E-Periodica comprising real-world magazines with complex document structures for evaluation. Our results demonstrate that our DSG outperforms commercial OCR tools and, on top of that, achieves state-of-the-art performance. To the best of our knowledge, our DSG system is the first end-to-end trainable system for hierarchical document parsing.

The construction of a character dataset is an important part of the research on document analysis and recognition of historical Tibetan documents. The results of character segmentation research in the previous stage are presented by coloring the characters with different color values. On this basis, the characters are annotated, and the character images corresponding to the annotation are extracted to construct a character dataset. The construction of a character dataset is carried out as follows: (1) text annotation of segmented characters is performed; (2) the character image is extracted from the character block based on the real position information; (3) according to the class of annotated text, the extracted character images are classified to construct a preliminary character dataset; (4) data augmentation is used to solve the imbalance of classes and samples in the preliminary dataset; (5) research on character recognition based on the constructed dataset is performed. The experimental results show that under low-resource conditions, this paper solves the challenges in the construction of a historical Uchen Tibetan document character dataset and constructs a 610-class character dataset. This dataset lays the foundation for the character recognition of historical Tibetan documents and provides a reference for the construction of relevant document datasets.

Tibetan is a low-resource language, and Tibetan culture carried by historical Tibetan documents is an important part of Chinese civilization. The study of historical Tibetan documents is of great significance to the protection of Tibetan culture and the promotion of Chinese culture. Character segmentation is an important step in image analysis and recognition of historical Tibetan documents. However, the following three challenges prevent solving problems of character segmentation in historical Tibetan documents: 1) the text lines have different degrees of tilt and twist; 2) there are many complex situations such as overlapping, crossing, touching and breaking character strokes; and 3) these documents are written by different people with different stroke styles. To resolve these problems, we propose a character segmentation method based on key feature information for historical Tibetan documents. The proposed method consists of three parts: 1) projection and syllable point location information are used to shorten the text lines of historical Tibetan documents and establish a character block database; 2) the local baseline of the character block is detected by using the location information of syllable points or combined with horizontal projection and straight line detection, and the character block is divided into two areas above and below the baseline, and different segmentation methods are adopted; and 3) in view of the large difference in stroke styles, three stroke attribution distances are used to complete the attribution. The experimental results show that the method proposed in this paper can effectively solve the problem of character segmentation of historical Tibetan documents and achieve a better character segmentation effect, which also provides a reference for the relevant document character segmentation.