DLIA99 Discussion summary:

 

IV: Extraction and learning of spatial relation

Discussion chair: Chin Suen (Concordia University)

Three papers were presented in this session, i.e.

1. "Learning regions of interest in postal automation" by H. Walischewski

   Talks about digital representation of mail pieces, and learning models
   to find the regions of interest such as the target address. The average
   recognition rate is about 92%.

2. "Interval-Algebra based block layout analysis and document template
   generation" by R. Singh et al

   Interval-algebra is introduced to capture the block layout descriptions
   in qualitative description of the entire class of documents.

3. "On the application of Voronoi diagrams to page segmentation" by
   K. Kise et al

   Aims at the development of general representations of the physical
   structure of document images by creating a point Voronoi diagram, then
   an area Voronoi diagram of connected components, and finally a neighbor
   graph.

In the discussions, we talked about methods of finding the interested
items in the image document and established the fact that they are
application dependent. Hence the characteristics of the items as well
as their environment must first be established. Then physical measurements
such as size and distances can be applied afterwards.

In addressing training and learning methods, both supervised and semi-
supervised techniques have been discussed. A large number of samples are
required to derive the parameters and thresholds to minimize errors.

Probabilistic techniques and statistics also entered into the discussions
as well as the use of ranking.