MSc. thesis projects at AI Institute Groningen University.

Research Programme: Autonomous Perceptive Systems

• Camera-based text detection and reading
  A set of digital photographs in the city environment as well as in our robocup-field environment has been collected, containing hundreds of text images of advertisements, road signs, and other text patterns. This material must be annotated manually, which is a tedious process. Methods need to be developed to speed up the labeling (truthing) process by using advanced text detection and classification methods to support the annotator.
  

  Examples of texts in the man-made physical environment. Do you think it is easy to "OCR" these? Forget it. Humans still outperform the machine.

• Multi-sensory biologically plausible robot navigation
  Recent projects have been performed using self-organized Kohonen landmark maps. Since odometry is miserably biased and erroneous, only the linearly traveled distance between landmarks is available. It appears that principal-components analysis on the distance matrix contains the information for a map reconstruction. If this basic information is enhanced using behavioral heuristics (navigational tuning), the accuracy of such a 'cognitive' map may be enhanced considerably. Animal behavior contains a large amount of such heuristics. As an example: if I see landmark Q in front, what do I see at that point on the map when I look to the left (landmark environment 'Q-left') and to the right (landmark environment 'Q-right'). Even insects display such navigation-support behaviors. The challenge is to create robust mechanisms that will help out in real robot systems if other, technical, means of navigation such as GPS fail.
  

  A self-organized landmark map of a robocup field with normalized color images and the sonar vectors

Perceptive Systems

• Handwriting recognition in historical manuscripts
  Preserving the cultural heritage requires digitization and annotation. A huge problem concerns the existance of thousands of documents which are available as scans, but it is impossible to use search engines for the handwritten content. In our lab new methods are being developed, in cooperation with the Dutch National Archive and other institutions. Machine-learning and data-mining approaches are used to exploit the ever-increasing available computing power.
  

  A fragment of a historical manuscript with difficult foreground/background separation and typical character shapes

• Automatic Writer Identification
  Recently, advances have been made in our lab in the area of automatic writer identification on the basis of off-line scanned handwritten paragraphs of text. A number of research questions remain, in the area of the scalability of the approach. Today, tests are in the order of hundreds of writers. The goals is to expand this to thousands of writers. Data collection, annotation and classifier testing are required. New methods for classifier combination must be developed. A software environment consisting of java, C(++) and Linux scripts exists.

  A questioned sample (top, orange border) and a hit list of the forensic writer search engine. The nearest neigbour is the correct writer (image border marked orange).