Pattern recognition process

Details Pattern recognition process Pattern recognition process

2000-03-01

2004-02-17

Boudreau, Leo (Department: 2621)

Image analysis

Pattern recognition

Reexamination Certificate

active

06694054

ABSTRACT:

FIELD OF THE INVENTION
The present invention relates to a pattern recognition process independent of the pattern orientation and size.
The invention can be applied to digital image processing and, more particularly, in the field of document recognition, on technical documents, for example.
BACKGROUND OF THE INVENTION
Two types of orientation-independent recognition methods are known:
methods consisting of a preliminary calculation of the pattern orientation and its correction to bring it to a reference position,
non-reorientation pattern processing methods.
Pattern correction-based pattern recognition methods induce deformations of the pattern. These deformations may be sufficiently significant to penalise the classification of the pattern, irrespective of the method used (K-NN method (K-nearest neighbour), neuronal network). For example, in the case of optical character recognition, this results in the production of incorrect characters.
This phenomenon is not critical if the recognition of a structured text is required, since a structured text generally has a fixed orientation and a context represented by paragraphs, words, a reduced number of character fonts and any other structural information. This context represents a significant volume of preliminary information which can help resolve recognition ambiguities.
In the case of technical drawings or maps, the orientation of the text, corresponding to annotations and place names, is very variable. The context cannot be generalised for the entire document, since the formalism of the character strings depends on the object described. Several fonts are used and the character size and boldness are subject to variation. Consequently, while a context does exist in such documents, it can only be local, i.e. limited to a restricted area of the drawing.
Non-reorientation pattern processing methods describe the pattern using invariant specific descriptors with reference to the rotation, translation or homothetic transformations. To produce this description, these methods require either the extraction of structural primitives or the use of classification techniques rendering the problem invariant with reference to the above-mentioned transformations.
The main methods known use either invariant moment extraction, circular primitive extraction or the extraction of Fourier descriptors on the outlines of the pattern to be recognised.
Invariant moment extraction requires the use of moments of very high orders to describe a pattern correctly. This progression in the orders of moments may pose classification problems due to the dynamics of the calculated values and error propagation.
Consequently, invariant moment extraction from a pattern generates highly complex algorithms. This has a considerable penalising effect on the pattern recognition process processing speed.
Circular primitive extraction consists of describing a pattern by surrounding it with a circle from which a probe is sent to the centre of the circle. This probe is used to describe the spectral profile of the pattern according to the radius of the circle. By defining a sufficient number of radii, for each pattern, it is then possible to constitute a table of characteristic values of the pattern. In order to determine the pattern, the values in the table are compared to reference values.
Circular primitive extraction also generates highly complex algorithms. For example, if the patterns to be recognised correspond to symbols, the use of circular primitives may generate tables comprising a very large number of values due to the need, in this case, to define a large number of radii for the different probes.
In addition, these methods pose problems related to noise-sensitivity on the image (signal
oise ratio).
The extraction of Fourier descriptors on the outlines of the pattern poses a large number of problems.
Firstly, it is difficult, or even impossible, to determine invariant primitives in relation to changes of scale and orientation which are reliable and relevant. Secondly, the descriptors are not complete (enabling the reconstruction of the image of the pattern from descriptors), which conveys the weakness of the description. Finally, experiments demonstrate that it is difficult to determine a relevant order from which patterns are described perfectly, while remaining insensitive to calculation noise.
SUMMARY OF THE INVENTION
The invention does not have these disadvantages. It relates to a pattern recognition process successively comprising a pattern detection step, a characterisation step of the pattern obtained in the detection step and a recognition step of the pattern obtained in the characterisation step. The pattern characterisation step is a step used to obtain invariant descriptors of the pattern D
f
(p,q) in the Fourier-Mellin space such that:
D
F
(
p,q
)=
M
F
(
p,q
)[
M
F
(0,1)]
−q
[M
F
(0,1)]
q
[M
F
(0,0)]
p
where M
F
(p,q) is the Fourier-Mellin transform of the pattern, with p and q coefficients conveying the invariance of the pattern in relation to rotation and change of scale, respectively.
The pattern recognition process according to the invention can be applied to all pattern types. For example, a pattern may be finite set of points defining a curve (closed or not) and/or a surface.
An advantage of the invention is that it enables pattern recognition by disregarding the overall context of the pattern and not applying geometric transformations to the pattern.
In many scenes, patterns may be attached together or overlap with each other. The recognition of these patterns represents a difficult problem to solve. The process according to the invention provides an effective solution to this problem.


REFERENCES:
patent: 5465308 (1995-11-01), Hutcheson et al.
patent: 5561718 (1996-10-01), Trew et al.
patent: 6282300 (2001-08-01), Bloom et al.
“Experiments on Pattern Recognition Using Invariant Fourier-Mellin Descriptors”. Sheng et. al, Jounal of the Optical Society of America vol. 3, No. 6, Jun. 1986, pp. 771-776.*
S. L. Diab, et al., Automatic Target Recognition VII, vol. 3069, pp. 269-280, “Scale and Translation Invariant Detection of Targets Varying in Fine Details”, Apr. 22-24, 1997.
Q. Chen, et al., IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 16, No. 12, pp. 1156-1168, “Symmetric Phase-Only Matched Filtering of Fourier-Mellin Transforms for Image Registration and Recognition”, Dec. 1994.

Affiliated with

Also associated with

Rating

Pattern recognition process does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Pattern recognition process, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Pattern recognition process will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-3352026