RR_03_02

assemblies to lower-level assemblies via 'bubbles' that
are labeled. Each labeled bubble shows a single assembly.
Accordingly, there is a tremendous need for layout
recognizers that can accurately segment the drawings into
their individual assemblies and automatically set up
hyper-links among the assemblies of different levels.
Figure 6: Parts pages require segmentation
In [2], we discussed our research on recognizing the
layout of parts pages in vector format. Applying
geometric proximity algorithms we achieved accuracies
on the order of 84% for grouping art into the proper
segments. Incorporating heuristics regarding the location
of detail labels, the recognition of leader lines and
attaching text to those leader lines increased our layout
recognition accuracy to -97%. Failures generally occur
when parts of a detail are deliberately offset from the
main body to show separation within the detail or when
there is considerable space between the body of the detail
and an orientation arrow (Figure 7). To achieve
accuracies much closer to 100% will require recognition
of the objects being presented, re-mapping the technical
illustration to the real-world object.
4. Raster vs. vector considerations
Layout recognition research has predominantly
focused on the understanding of the layout of raster
3
images. While some of the examples we have discussed
do occur as
Figure 7: Orientation arrows can cause
segmentation errors
raster drawings, most are vector images produced in
modern illustration tools. Vector graphics are generally
easier to analyze than raster - there is no need for optical
character recognition, for example. However vector
graphics present unique challenges that are not present in
raster images and require a very different approach. For
example, techniques involving histograms or ink density
translate poorly, if at all, to vector diagrams. We need
new recognition algorithms based on the rendering
characteristics of vector graphics. This is a fertile area for
new research.
5. Summary
Aerospace companies depend heavily on a wide
variety of technical documentation in tremendous
quantities. We have presented just a few of the types of
layout recognitIOn challenges found in technical
documentation. Solutions must be highly scalable and
extremely accurate. We hope to inspire new research in
this relatively unexplored area of document recognition.
6. References
[1]707,727-777 Standard Wiring Practices Manual, 06-
54446, The Boeing Company
[2] L.S . Baum, J.H.Boose, and R.I. Kelley, "Graphics
Recognition for a Large-Scale Airplane Information
System", Graphics Recognition, Algorithms and Systems,
Second International Workshop, GREC '97, Springer,
Nancy, France, August 1997, pp. 291-30 I.