Using Multi-level Segmentation Features for Document Image Classification

Document Image classification is a crucial step in the processing pipeline for many purposes (e.g. indexing, OCR, keyword spotting) and is being applied at early stages. At this point, textual information about the document (OCR) is usually not available and additional features are required in order to achieve higher recognition accuracy. On the other hand, one may have reliable segmentation information (e.g. text block, paragraph, line, word, symbol segmentation results), extracted also at pre-processing stages. In this paper, visual features are fused with segmentation analysis results in a novel integrated workflow and end-to-end training can be easily applied. Significant improvements on popular datasets (Tobacco-3482 and RVL-CDIP) are presented, when compared to state-of-the-art methodologies which consider visual features.