Logical structure extraction from scanned documents

Abstract

Logical structure extraction from various documents has been a longstanding research topic because of its high influence on a wide range of practical applications. A huge variety of different types of documents and, as a consequence, the variety of possible document structures make this task particularly difficult. The purpose of this work is to show one of the ways to represent and extract the structure of documents of a special type. We consider scanned documents without a text layer. This means that the text in such documents cannot be selected or copied. Moreover, you cannot search for the content of such documents. However, a huge number of scanned documents exist that one needs to work with. Understanding the information in such documents may be useful for their analysis, e. g. for the effective search within documents, navigation and summarization. To cope with a large collection of documents the task should be performed automatically. The paper describes the pipeline for scanned documents processing. The method is based on the multiclass classification of document lines. The set of classes include textual lines, headers and lists. Firstly, text and bounding boxes for document lines are extracted using OCR methods, then different features are generated for each line, which are the input of the classifier. We also made available dataset of documents, which includes bounding boxes and labels for each document line; evaluated the effectiveness of our approach using this dataset and described the possible future work in the field of document processing.