TACTFUL: A Framework for Targeted Active Learning for Document Analysis.

Document Layout Parsing is an important step in an OCR pipeline, and several research attempts toward supervised, and semi-supervised deep learning methods are proposed for accurately identifying the complex structure of a document. These deep models require a large amount of data to get promising results. Creating such data requires considerable effort and annotation costs. To minimize both cost and effort, Active learning (AL) approaches are proposed. We propose a framework TACTFUL for Targeted Active Learning for Document Layout Analysis. Our contributions include (i) a framework that makes effective use of the AL paradigm and Submodular Mutual Information (SMI) functions to tackle object-level class imbalance, given a very small set of labeled data. (ii) an approach that decouples object detection from feature selection for subset selection that improves the targeted selection by a considerable margin against the current state-of-the-art and is computationally effective. (iii) A new dataset for legacy Sanskrit books on which we demonstrate the effectiveness of our approach, in addition to reporting improvements over state-of-the-art approaches on other benchmark datasets.