Robust Character Recognition in Low-Resolution Images and Videos

Kopf, Stephan ; Haenselmann, Thomas ; Effelsberg, Wolfgang

Kopf2004c.pdf - Published

Download (399kB)

URN: urn:nbn:de:bsz:180-madoc-9807
Document Type: Working paper
Year of publication: 2005
Publication language: English
Institution: School of Business Informatics and Mathematics > Sonstige - Fakultät für Mathematik und Informatik
MADOC publication series: Veröffentlichungen der Fakultät für Mathematik und Informatik > Institut für Informatik > Technical Reports
Subject: 004 Computer science, internet
Subject headings (SWD): Bildsegmentierung , Bildanalyse , Filmanalyse , Optische Zeichenerkennung
Individual keywords (German): OCR , Krümmungsbasierter Skalenraum , CSS
Keywords (English): OCR , Curvature scale space , CSS , Character segmentation
Abstract: Although OCR techniques work very reliably for high-resolution documents, the recognition of superimposed text in low-resolution images or videos with a complex background is still a challenge. Three major parts characterize our system for recognition of superimposed text in images and videos: localization of text regions, segmentation (binarization) of characters, and recognition. We use standard approaches to locate text regions and focus in this paper on the last two steps. Many approaches (e.g., projection profiles, k-mean clustering) do not work very well for separating characters with very small font sizes. We apply in a vertical direction a shortest-path algorithm to separate the characters in a text line. The recognition of characters is based on the curvature scale space (CSS) approach which smoothes the contour of a character with a Gaussian kernel and tracks its inflection points. A major drawback of the CSS method is its poor representation of convex segments: Convex objects cannot be represented at all due to missing inflection points. We have extended the CSS approach to generate feature points for concave and convex segments of a contour. This generic approach is not only applicable to text characters but to arbitrary objects as well. In the experimental results, we compare our approach against a pattern matching algorithm, two classification algorithms based on contour analysis, and a commercial OCR system. The overall recognition results are good enough even for the indexing of low resolution images and videos.
Additional information:

Das Dokument wird vom Publikationsserver der Universitätsbibliothek Mannheim bereitgestellt.

Metadata export


+ Search Authors in

+ Download Statistics

Downloads per month over past year

View more statistics

You have found an error? Please let us know about your desired correction here: E-Mail

Actions (login required)

Show item Show item