Team Members: Thomas C. Henderson
Week
1:
The semantic interpretation
of industrial drawings is a very difficult
problem. Tombre (Tombre 1998a) summarizes the state of the field for CAD
drawings and as a document image analysis problem (Tombre
1998b, Tombre 1998c}.
An earlier overview by Haralick and Kanungo (Kanungo 1995} gives
insight into the scale of the technical drawing problem (e.g., about 250
million drawings are generated annually), as well as a discussion of
performance measurement. A major problem
pointed out by these reviewers is that there is a
dearth of work relating high-level models of documents, drawings or graphics to
the various levels of analysis.
The goal of technical
drawing analysis is to interpret the contents of
an image, such as that shown
here (part of an engineering drawing):

Several systems have been
developed along these lines, although none works well in an automated
fashion. Various problems arise; e.g., when
text touches graphics or when there is noise in the scanned image or when the
thresholds of the image-analysis codes support only a few classes of
drawings. One notable system is
CELESSTIN (Ahsoon 1995, Tombre
1993a}; however, as pointed out by Haralick,
CELESSTIN suffers from the fact that it is based on a hierarchical rule system
and has many rules; furthermore, whatever interpretive models the rules
instantiate exist only in the thresholds and logic of the hand-coded rules.
In the standard form of
analysis, the scanned drawing is digitized, noise is removed, text and graphics
are recognized, graphics is vectorized, dimensions
are extracted and the whole drawing is analyzed by applying knowledge rules.
Each of these steps normally uses just a single set of thresholds. Our goal is to allow a wide range of
thresholds to generate a potentially large search space (hopefully including
the correct components), and then to apply domain constraints to prune the
space.
The interpretation of engineering drawings
requires:
Thus, we will develop models based on these
issues. Models include:
Week 1: Define Project
Week 2: Develop physical degradation model
Develop image formation model
Week 3: Explore use of color in models
Week 4: Develop line segment model
Develop image smoothing approach
Explore use of linear (and
nonlinear) filters
Week 5: Explore use of gradient, 1st
derivative, etc.
Explore line segment detector
alternatives
Week 6: Define performance measures
Develop benchmark dataset and ground truth
Week 7: Implement first cut image analysis system
Week 8: Measure performance of image analysis
Week 9: Explore various structural methods
Week 10:
Compare various structural methods
Week 11:
Compare structural methods
Week 12:
Explore complete system sensitivity analysis
Week 13:
Thanksgiving
Week 14:
Measure system performance
Week 15:
Optimize system and draw conclusions
Week 16:
Final results and report
Tombre 1998a "Analysis of Engineering Drawings:
State of Art and Challenges," Karl Tombre, Graphics
Recognition - Algorithms and Systems, Lecture Notes in Computer Science, Vol.
1389, Springer-Verlag, Berlin, April, 1998, pp
257-264.
Tombre 1993a "Don't Tell Mom I'm Doing Document
Analysis; She Believes I'm in the Computer Vision Field," Suzanne Collin,
Karl Tombre and Pascal Vaxiviere,
Proceedings of 2nd International Conference on Document Analysis and
Recognition, Tsukuba Science City (Japan), October, 1993, pp. 619-622.
Ahsoon 1995 "A Step Towards
Reconstruction of 3-D CAD Models from Engineering Drawings," Christian
Ah-Soon and Karl Tombre, Proceedings 3rd International
Conference on Document Analysis and Recognition", Montréal (
Tombre 1998b "Ten Years of Research in the
Analysis of Graphics
Documents: Achievements and Open Problems," Karl Tombre, Proceedings of the 10th Portuguese Conference on
Pattern Recognition,
}
Tombre 1998c Graphics Documents: Achievements and Open
Problems," Karl Tombre, Proceedings of the 10th
Portuguese Conference on Pattern Recognition,
Kanungo 1995 "Understanding Engineering Drawings:
A Survey," Tapas Kanungo,
Robert M. Haralick and Dov Dori, Proceedings of First IARP Workshop on Graphics
Recognition, University Park, PA, pp. 217-228, 1995.
Week
2: January 15 –
Physical Degradation
Model:
Physical paper engineering drawings can undergo several types of degradation:
Tearing:
A rip will be characterized as a straight line segment starting within some distance of an edge and running at some angle straight to the edge of the image with the following parameters:
:
A hole will be characterized as a circle with the following parameters:
A removed area will be characterized by the following parameters:
· shape: selected from a predefined set of shapes defined as pixel sets
· scale: selected from a normal distribution over N(mu,sigma)
Staining:
Staining will be parameterized by:
[Note: this model is identical to the removed area model.]
Folding:
Folding will be modeled as a line through the image and will be characterized by the following parameters:
Image Formation Model
The engineering drawing image is produced by a linear camera array which moves along the x-axis (i.e., along the columns) to scan the image. The following parameters are used to model the process: