Recognizing Figure Labels in Patents.

Ming Gong, Xin Wei, Diane Oyen, Jian Wu, Martin Gryder, Liping YangODU 2021

Abstract

Scientific documents often contain significant information in figures. The United States Patent and Trademark Office (USPTO) awards thousands of patents each week, with each patent containing on the order of a dozen figures. The information conveyed by these figures typically include a drawing or diagram, a label, caption and reference text within the document. Yet associating the short bits of text to the figure is challenging when labels are embedded within the figure, as they typically are in patents. Using patents as a testbench, this paper highlights an open challenge in analyzing all of the information presented in scientific/technical documents - namely, there is a technological gap in recognizing characters embedded in drawings, which leads to difficulties in processing the text associated with scientific figures. We demonstrate that automatically reading the figure label in patent diagram figures is an open challenge, as we evaluate several state-of-the-art optical character recognition (OCR) methods on recent patents. Because the visual characteristics of drawings/diagrams are quite similar to that of text (high contrast, width of strokes, etc), separating the diagram from the text is challenging and leads to both (a) false detection of characters from pixels that are not text and (b) missed text that is critical for identifying the figure number. We develop a method for automatically reading the patent figure labels by first identifying the bounding box containing the label using a novel non-convex hull approach, and then demonstrate the success of OCR when the text is isolated from the diagram.
BibTeX
@inproceedings{gong2021recognizing,
  author = {Ming Gong and Xin Wei and Diane Oyen and Jian Wu and Martin Gryder and Liping Yang},
  title = {{Recognizing Figure Labels in Patents.}},
  booktitle = {ODU},
  year = {2021},
}

Loading related content...

Figure