CalliScan

Abstract

In this work, a solution for handwriting text extraction from images with visual user assistance is proposed. Use of end-to-end systems that pipe together text detection and recognition is often awkward because the user cannot influence the detection stage. On the other hand, glossing over the word's regions to help system with text localization requires a manual job and can be unacceptable. This paper proposes a solution that gives visual cues to the user during a detection stage. These hints differ from traditional bounding boxes in two ways. Firstly, the found text is surrounded with polygonal bounding reflecting a possible complex nature of text blocks. Secondly, TextRadar scanning effect provides a non-overloaded camera view, helping the user to capture the most relevant part of the text on image on-the-fly. CalliScan works on-device and keeps the user's privacy. The evaluation study has shown that users need such a solution, but it is necessary to carefully handle the text layout complexity.