Note: Descriptions are shown in the official language in which they were submitted.
CA 02547304 1995-04-12
August 24, 1994 - 1015-1611
LUMEER DEFECT SCANNING INCLUDING MULTI-DIMENSIONAL
PATTERN RECOGNITION
BACKGROUND OF THE INVENTION
The present invention relates generally to wood product
processing, and particularly to an automated method and apparatus
for detection of defects and other abnormal conditions on the
surface of lumber or other solid wood products.
Automatic detection of defects and abnormal conditions
improves wood processing operations. Overall production efficiency
and product quality increases by automatic defect detection and
corresponding product grading, processing or remedial action.
Unfortunately, many existing defect scanning techniques have been
complex and not always suitable across a sufficiently broad set of
defect types.
Attempts to apply image processing technology have generally
?0 failed to adequately identify defects in a wood article. In
particular, image processing algorithms generally suffer from an
inability to find or measure the size of defects and also suffer
from misclassification, i.e., sometimes identifying a normal grain
area inappropriately as a defective area. Typically, image
?5 processing algorithms suffer from a lack of data representative of
wood cell structure. Image processing algorithms receive a massive
1
CA 02547304 1995-04-12
volume of image-based data, i.e., data representing the intensity
of color in various color bands at specific locations on a wood
grain article. Unfortunately, such massive image-based data does
not directly relate to the wood grain structures of interest.
Accordingly, while image processing' algorithms receive and process
massive amounts of data, the algorithms suffer from lack of real
information concerning the wood grain structures to be
characterized. Thus, where image processing algorithms can, at
best, emulate vision and analyze the resulting image, such
emulation and analysis typically fails to accurately classify grain
structures, i.e., differentiate between normal grain and abnormal
or defective grain.
It would be desirable, therefore, to improve automated defect
detection and reporting apparatus and methods to more accurately
characterize wood grain conditions throughout a broad set of
potential defects and wood grain structure conditions.
SUMMARY OF THE INVENTION
?0 In accordance with the present invention, a method for
identifying a defect relative to a given wood article begins by
collecting a first set of data values characterizing a portion of
a sample wood article surface, the portion of the sample wood
article surface corresponding to a known defect. The data values
;5 collected include at least one value responsive to wood grain
structure at the portion of said wood article. At least one
2
CA 02547304 1995-04-12
mathematic function is then selected for application to the first
set of data values, the mathematic function resulting in a given
clustering of or distance between said data values according to
multivariate pattern recognition analysis. The first set of data
values and selected mathematic function thereby establish a
training set. The next step employed under the present invention
calls for collection of a second set of data values characterizing
a portion of said given wood article. The second set of data
values corresponds to the first set of data values and is also
applicable the selected at least one mathematic function to
indicate a clustering of or distance between the second set of data
values. If the selected mathematic function indicates sufficiently
minimum distance between the second set of data values at or below
the given distance, then the portion of the given wood article is
identified as corresponding to the known defect associated with the
first set of data values.
In accordance with one aspect of the present invention, a
method for detecting the presence of at least one of a set of wood
cell structure conditions at a wood article surface includes a
preliminary step of executing a training session wherein for each
of said set of wood cell structure conditions at least one training
vector is defined. Each training vector includes a designation of
the associated wood cell structure condition and a set of training
z5 vector data values characterizing the associated wood cell
structure condition. The method next includes executing a scanning
3
CA 02547304 1995-04-12
71208-89
procedure against inspection points of said given wood article
wherein a scan vector is collected for each inspection point. Each
scan vector includes location data indicating a physical location
for the associated inspection point and includes a set of scan
vector data values corresponding to said training vector data
values. Under multivariate pattern recognition analysis, each scan
vector is applied to a mathematical function associated with each
training vector, i.e., a function sufficiently clustering or
minimizing distance between data points of the training set vector.
If the function causes similar clustering for a scan vector, then
the location data associated with the scan vector may be further
associated with the wood grain condition of the statistically
corresponding training vector.
The present invention further provides a method identifying
the condition of inspection points of a wood article surface by
first scanning a plurality of inspection points of the wood article
surface. The scanning step includes collection f or each inspection
point both reflective grain defect data and tracheid effect data.
The process continues by identifying a first class of surface
conditions by reference to the grain defect scanning data, and
identifying a second class of surface conditions by reference to
the tracheid effect data while excluding from consideration those
inspection points identified as being said first class surface
conditions.
4
CA 02547304 1995-04-12
71208-89
In accordance with another aspect of the present
invention, there is provided a method for identifying a
defect relative to a given wood article, said method
comprising the steps: collecting a first set of sensor data
values characterizing a portion of a sample wood article
surface, said portion of said sample wood article surface
corresponding to a known defect, said data values including
at least one value representing wood grain structure at said
portion of said wood article; selecting at least one
mathematic function applied to said set of data values and
providing a mathematical characterization of said set of
data values; collecting a second set of data values
characterizing a portion of said given wood article, said
second set of data values corresponding to said first set of
data values; applying said mathematic function to said
second set of data values; and if the resulting mathematical
characterization of said second set of data values
corresponds to the first set of data values then identifying
said portion of said given wood article as corresponding to
said known defect.
In accordance with yet another aspect of the
present invention, there is provided a method for detecting
presence of at least one of a set of wood cell structure
conditions at a wood article surface, said method comprising
the steps: executing a training session wherein for each
member of said set of wood cell structure conditions at
least one training vector is defined, each training vector
including a designation of the associated wood cell
structure condition and a set of training vector data values
characterizing the associated wood structure condition, said
training vector data values including at least one
representation of wood cell structure; executing a scanning
procedure against inspection points of said given wood
4a
CA 02547304 1995-04-12
71208-89D
article wherein a scan vector is collected for each
inspection point, each scan vector including location data
indicating a physical location for the associated inspection
point, each scan vector including location data indicating a
physical location for the associated inspection point and
including a set of scan vector data values corresponding to
said training vector data values; evaluating by pattern
recognition analysis each scan vector with each training
vector; and detecting presence of a given wood structure
condition at a given location when a training set vector
corresponding to said given wood structure condition
substantially matches under said pattern recognition
analysis a scan vector for said given location.
In accordance with yet another aspect of the
present invention, there is provided a method of identifying
the condition of inspection points of a wood article
surface, the method comprising the steps: scanning a
plurality of inspection points of said wood article surface,
said scanning step collecting for each inspection point both
reflective grain defect data and tracheid effect data;
identifying a first class of surface conditions by reference
to said grain defect scanning data; and identifying a second
class of surface conditions by reference to said tracheid
effect data while excluding from consideration those
inspection points identified as being said first class of
surface conditions.
In accordance with yet another aspect of the
present invention, there is provided a method of identifying
the condition of inspection points of a wood article
surface, the method comprising the steps: scanning a
plurality of inspection points of said wood article surface,
said scanning step collecting for each inspection point both
reflective grain defect data and tracheid effect data;
4b
CA 02547304 1995-04-12
71208-89D
identifying a first class of surface conditions by reference
to said grain defect scanning data; and identifying a second
class of surface conditions by reference to said tracheid
effect data while excluding from consideration those
inspection points identified as being said first class of
surface conditions.
4c
CA 02547304 1995-04-12
The subject matter of the present invention is particularly
pointed out and distinctly claimed in the concluding portion of
this specification. However, both the organization and method of
operation of the invention, together with further advantages and
objects thereof, may best be understood by reference to the
following description taken with the accompanying drawings wherein
like reference characters refer to like elements.
BRIEF DESCRIPTION OF THE DRAWINGS
l0 For a better understanding of the invention, and to show how
the same may be carried into effect, reference will now be made, by
way of example, to the accompanying drawings in which:
FIG. 1 illustrates a scanning apparatus collecting multi-
dimensional information relative to a wood article and including
information representative of article grain structure and of
article image.
FIG. 2 illustrates a side view of the apparatus of FIG. 1 as
?0 taken along lines 2-2 of FIG. 1.
FIG. 3 illustrates a scan vector of multi-dimensional data
collected by the apparatus of FIG. 1 relative to a specific
inspection point of a wood article.
;5
FIG. 4 illustrates a data representation of a wood article as
5
CA 02547304 1995-04-12
provided by the apparatus of FIG. 1 and comprising a collection of
scan vectors as shown in FIG. 3.
FIG. 5 illustrates a first use of the data representation
portrayed in FIG. 4 wherein certain detector outputs identify
defects best identified by such data and remaining detector outputs
identify defects best identified by remaining detector data.
FIG. 6 illustrates a training set vector of generally similar
structure to that of the scan vector of FIG. 3, but holding data
and mathematic functions characterizing a known grain structure or
defect as developed in a training session in anticipation of a
pattern recognition analysis.
L5 FIG. 7 illustrates application of the data representation of
FIG. 4 and a collection of training set vectors as shown in FIG. 5
to a pattern recognition process providing as output
characterization of grain structures at specific locations of a
wood article.
?0
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT
The present invention operates generally by point-to-point
inspection of a wood article while collecting and storing a variety
of information characterizing the wood article at each inspection
'5 point. Thus, for each inspection point on a given article a set or
vector of mufti-dimensional data is collected and stored. As may
6
CA 02547304 1995-04-12
be appreciated, two spatial dimensions locate the inspection point
on the wood article, i.e., X and Y locations relative to the wood
article. Additional dimensions of data stored in association with
each inspection point include information representing the
structure of wood grain at the inspection point. Other measurement
information, e.g., color and height, relative to the inspection
point further augment mufti-dimensional characterization of the
inspection point.
LO By modeling the inspection point as a vector of such multi-
dimensional information, multivariate pattern recognition
techniques are applied to characterize the inspection point. In
particular, information obtained at a given inspection point, i.e.,
a scan vector, is analyzed under multivariate analysis relative to
_5 data vectors of similar structure from a training set, i.e., a
training set vector, to statistically correlate each scan vector
with a defect or with clear wood. Because information provided in
the scan vector relates directly to the structure of wood grain at
the inspection point, methods of pattern recognition become more
0 useful, i.e., more useful relative to use of such pattern
recognition techniques using only imaged-based information, e.g.,
using only color of an inspection point. When used in conjunction
with distinct defect scanning and detection techniques, each
contributing to one or more dimensions of data, the scanning method
5 of the present invention reliably and automatically scans a wood
article to accurately detect and locate defect conditions therein.
7
CA 02547304 1995-04-12
The key to using multivariate pattern analysis is use of
mufti-dimensional data, i.e., a number of channels of data measured
at essentially the same inspection point and at essentially the
same time. Wide variation in response among the various channels
for various defect conditions improves defect differentiation under
pattern recognition techniques. The variation is not necessarily
of one data channel relative to another data channel, but a scan
vector viewed as a whole under multivariate analysis varies widely
for one- condition, e.g. a knot, relative to another condition,
e.g., clear wood. Collecting relative to a given inspection point
a variety of data channels, including both wood cell structure
responsive signals and image responsive signals, improves wood
grain condition distinction under pattern recognition analysis.
Mufti-dimensional data values taken from a scan vector are
applied to a function previously found to sufficiently cluster
reference data, i.e., sufficiently minimize distance between data
points in n-dimensional space for a training set vector obtained as
reference data by measurement of a known grain structure condition.
For example, a pre-defined training set vector for a "diving grain
knot" associates a descriptor representing a "diving grain knot"
with a collection of mufti-dimensional data found to coincide with
such defect. A function is then selected which clusters the data
points of the training set vector. When the mufti-dimensional data
of the scan vector is applied to this function and similar
clustering results, the inspection point associated with the scan
a
CA 02547304 1995-04-12
71208-89
vector is characterized according to the descriptor, e.g., diving
grain knot, of the training set vector.
The following description will show a simplified method of
obtaining a variety of information relative to an inspection point
to define a vector of multi-dimensional data characterizing that
portion of a wood article. In accordance with the present
invention, such multi-dimensional data includes data representative
of wood-grain structures at the inspection point, height of the
inspection point, and image-based information, e.g., representing
color intensity of the inspection point in various color bands.
Overall, a robust and widely varying collection of data
characterizes both the image and wood grain structure at each
inspection point thereby supporting successful application of such
multi-dimensional data to pattern recognition analysis.
FIGS. 1 and 2 illustrate an apparatus 20 collecting inspection
point information. A first source of grain structure related
information used under the present invention is obtained by
reflective grain defect scanning. The apparatus 20 and reflective
grain defect scanning technique illustrated in FIGS. 1 and 2 is
more fully described and illustrated in U.S. Patent No. 5,252,836,
9
CA 02547304 1995-04-12
71208-89
Generally, the method of grain detection and characterization
shown in U.S. Patent No. 5,252,836 shall be referred to herein as
"reflective grain defect scanning" and described generally as
follows. For a complete discussion of reflective grain defect
scanning including most desirable angles of orientation for
detectors 22 and 24 a more practical flying-spot implementation,
and methods of employing data obtained therefrom, U.S. Patent No.
5,252,836 should be consulted. Accordingly, only a general
description of reflective grain defect scanning need be presented
herein in support of the present invention.
Reflective grain defect scanning is particularly well adapted
for detection and precise size measurement of grain defects, e.g.,
knots. Data obtained from reflective grain defect scanning
provides not only the precise location of grain defects, but also
precise measurement of the size of such grain defects. As
discussed more fully hereafter, reflective grain defect scanning
may be employed to first identify those defects, e.g., knots, for
which it is best suited. Once this class of grain defects are
identified, located, and measured precisely in size, the remaining
portions of wood article 12 may be subjected to other methods of
analysis, ,e. g., methods best suited for detection of remaining
types of defects, e.g., defects other than knots. Reflective grain
defect scanning is also very sensitive to early wood/late wood
CA 02547304 1995-04-12
grain patterns and is therefore well suited for differentiating
flat and vertical wood grain patterns.
FIGS. 1 and 2 illustrate defect detection apparatus 20 using,
among other methods, reflective grain defect scanning. In FIGS. 1
and 2, an elongate wood grain article 12, having a longitudinal
axis 14 and inspection surface 16, is subject to defect scanning at
an inspection point 18 of surface 16. As used herein, the term
"inspection point" shall refer to a point of incidence of scanning
l0 light directed upon an inspection surface, not necessarily a static
point on the inspection surface 16. Article 12 has wood fiber
cells lying substantially along its longitudinal axis 14, but may
have grain defects or gross deviation from such normal direction of
the wood fiber cells, e.g., knot wood. Apparatus 20 identifies by
reflective grain defect scanning grain defects at the inspection
point 18 of article 12.
Translation of article 12 relative to apparatus 20 and along
longitudinal axis 14, e.g., as in longitudinal material feeding
applications, provides grain defect scanning along a line 18'
corresponding to a plurality of inspection points 18 as defined by
such longitudinal movement of article 12 relative to apparatus 20.
While an actual implementation of relative grain defect scanning
might employ a more complex scanner arrangement, e.g., flying-spot
laser scanning, the illustrations of FIG. 1 and 2 provide a
fundamental understanding of reflective grain defect scanning and
11
CA 02547304 1995-04-12
the nature of data obtained therefrom.
Apparatus 20 includes a pair of light detection devices,
detector 22 and detector 24, each bearing upon inspection point 18
and lying along the line of material feed, i.e., in the plane of
incidence 25 orthogonal to the surface 16 and containing the line
18'. As used herein, detector 22 shall be referred to as a
specular detector providing an output S. Detector 24 is a retro-
detector providing an output R. Detectors 22 and 24 lie generally
symmetrically within a plane of incidence 25. The lines of sight
to inspection point 18 for detection devices 22 and 24 are
substantially symmetric about a vertical reference axis 30 within
plane of incidence 25, normal to surface 16, and coincident with
point 18. A light source 26 directs a collimated light beam 27,
e.g., a low pcwer laser beam, toward point 18 and substantially, as
close as possible, along the line of sight between detector 24 and
inspection point 18. Each detector 22 and 24 produces outputs S
and R representing a level of reflected light energy detected. A
discrimination circuit 28 receives the outputs S and R from
detectors 22 and 24, respectively, for each inspection point 18
scanned.
As explained more fully in U.S. Patent No. 5,252,836, light
beam 27 reflects from surface 16 according to two models of
reflection, i.e., specular and/or diffuse reflection. Generally,
if inspection point 18 corresponds to a well behaved grain pattern,
12
CA 02547304 1995-04-12
then the specular model of light reflection dominates. If,
however, inspection point 18 corresponds to defective grain
structure, then the diffuse reflection model dominates. If
detectors 22 and 24 indicate reflected light of substantially the
same magnitude intensity, i.e., the ratio S/R approximately equal
to unity, then the diffuse light reflection model applies and the
inspection point 18 is taken to correspond to a wood grain defect.
If, however, the specular detector 22 receives more light than that
of retro-detector 24, i.e., the ratio S/R is greater than unity,
then the specular light reflection model dominates, indicating a
normal wood grain pattern. Comparing the ratio S/R to a threshold
value characterizes wood grain structures as normal or abnormal at
the inspection point 18.
Thus, the methods and apparatus of reflective grain defect
scanning provide information representing the structure of wood
grain at a given inspection point 18. Given an understanding of
the disclosure of U.S. Patent No. 5,252,836, it will be understood
that reflective grain defect scanning provides a stream of data
characterizing a wood article. In particular, for each wood
article scanned, a collection of samples are obtained. Each sample
includes a sequence of information beginning with an X value and a
Y value representing the position of the associated inspection
point 18 relative to the surface 16 of the wood article 12.
z5 Additional information stored in each sample includes an output S
magnitude, an output R magnitude, and a calculated S/R ratio. As
13
CA 02547304 1995-04-12
71208-89
used in the disclosure of U.S. Patent No. 5,252,836, the ratio S/R
is compared to a given threshold value to detect a ratio
substantially equal to unity, i.e., to discriminate between normal
and defective grain patterns. Under the present invention,
however, it is suggested that the magnitude of outputs S and R be
preserved for use in later pattern recognition algorithms.
Preserving the magnitude of detector outputs S and R as multi-
dimensional data applied to pattern recognition algorithms
increases data channel variation as a function of defect type as is
desirable in pattern recognition algorithms.
While use of reflective grain defect scanning has proven
successful in grain defect detection, the present invention
incorporates additional information into each inspection point
sample, i.e., additional data dimensions, as a function of other
information channels characterizing the inspection point 18.
One category of grain structure information collected relative
to inspection point 18 is based on the "tracheid effect." U.S.
Patent No. 3, 976, 384 filed January 16, 1975, issued August 24, 1976
and entitled METHOD AND APPARATUS FOR DETECTING TIMBER DEFECTS
describes the tracheid effect and its use in detecting wood grain
defects.
Generally, the tracheid effect results from transmission of
14
CA 02547304 1995-04-12
light within a wood article more freely along and within the cell
structure rather than transverse to the cell structure. The
tracheid effect is more obvious in softwoods, in the direction of
the long, hollow tracheid cells that comprise most of the softwood
structure. Light impinging a wood surface scatters into the cell
structures, but scatters more freely along the length of cells then
transverse to the cells. Light traveling transverse to the
longitudinal axis of the cells quickly encounters cell walls and
light transmission in this direction is restricted. As a result,
light scattering within a wood article having normal grain
structure produces a characteristic oval pattern lying along the
longitudinal axis of the cell structures. Wood grain of abncrmal
structure, however, tends to be uniformly diffuse from the point of
light entry into the wood article, and a characteristic circular
region surrounding the point of light entry results.
Measurement of the tracheid effect is accomplished by a
variety of mechanisms, U.S. Patent No. 3,976,384 showing one such
mechanism. Generally, by measuring the light intensity at a
location very near the point of light entry and along the normal
cell length indicates whether the characteristic oval pattern
exists or the characteristic circular pattern exists. More
particularly, given an average cell length of two to three
millimeters, the characteristic oval pattern should produce at a
?5 distance of approximately two to three millimeters from the point
of light entry and along the normal grain pattern a relatively
CA 02547304 1995-04-12
greater light intensity. Similarly, if the grain structure is
abnormal, then light intensity at a location approximately two to
three millimeters from the point of light entry and along the
normal grain pattern is relatively low. Thus, an apparatus
measuring the tracheid effect detects a magnitude of light
intensity at a location adjacent to, e.g., two to three
millimeters, and along the normal grain pattern from the point of
light entry into the wood article. By incorporating the
measurement of the tracheid effect into a collection of other data
l0 representing cell structure at a given inspection point, a more
comprehensive and robust characterization of the cell structure
results.
In FIGS . 1 and 2 , 1 fight detector 2 3 in a f first form represents
a high precision imaging optics light detector, directed generally
toward the inspection point 13, focusing with high precision at an
adjacent inspection point 19. More particularly, inspection point
19 lies along the longitudinal axis 14 of wood article 12 relative
to inspection point 18 at a distance corresponding to the length of
cell structures within the wood article 12, e.g., at a distance of
approximately two to three millimeters for softwoods.
As may be appreciated, such first form of light detector 23
must have :precise imaging optics capable of focusing at the
?5 inspection point 19, near the much brighter inspection point 18, to
accurately detect a magnitude of light intensity representative of
16
CA 02547304 1995-04-12
light transmitted through wood article 12. If light intensity at
inspection point 19 is of given magnitude, one assumes that such
intensity results from light transmission along and through normal
cell structures, i.e., cell structures at inspection point 18
oriented generally along the longitudinal axis 14 of wood article
12. Light intensity below this given magnitude, however, indicates
blockage of light transmission within wood article 12 and along
axis 14 implying an abnormal cell structure, i.e., not well
oriented along the longitudinal axis 14 of wood article 12.
Detector 23 in FIGS. 1 and 2 also represents an alternative
less complex and less expensive form, i.e., not requiring a high
precision imaging optics light detector. In particular, the
tracheid effect is useful in detecting presence of a variety of
wood surface conditions including small-area defects, e.g., knots,
and broad-area defects, e.g., stains, wood decay, and compression
wood. A high precision form of detector 23 is required for
detection and measurement of small-area defects, but a lower
resolution form of detector 23 will suffice for detection and
measurement of broad-area defects. By use of a mask in detector 23
obscuring the inspection point 18 and by use of a relatively lower
resolution and broader focusing light detection device, e.g., a
large-aperture photodiode, one detects and measures broad-area
defects by tracheid effect analysis.
As will explained more fully hereafter, by collecting in
17
CA 02547304 1995-04-12
conjunction with the tracheid effect measurement other wood grain
structure information, such as provided by the above-described
grain defect scanning method and apparatus, the less complex and
less expensive form of tracheid effect detector 23 may be employed.
More particularly, the above-described grain defect scanning method
and apparatus may be used to find and measure certain defects,
e.g., knots, and provide a basis for excluding data for such
portions of surface 16 from tracheid analysis. Excluding grain
defects-such as knots, leaves for analysis the broad-area defects
which may be adequately detected using the lower resolution, less
precise form of tracheid detector 23. Under such method, the
detector 23 output T is collected for each inspection point 18, but
when using the less precise form of tracheid effect detector 23
those inspection points 18 associated with certain defects, e.g.
knots, as indicated by the reflective grain defect scanning method
and apparatus, can be ignored when seeking broad-area defects based
on output T.
Thus, the tracheid effect is analyzed on a point-by-point
basis whereby for each inspection point 18, representing a point of
light entry into wood article 12, an adjacent inspection point 19
is viewed and a magnitude of light intensity measured thereat.
While the tracheid effect may be applied to a binary decision or
choice between normal wood and abnormal wood, i.e., by comparing
the measured intensity at inspection point 19 to a given threshold
light intensity differentiating normal wood from abnormal wood, the
18
CA 02547304 1995-04-12
magnitude of light intensity measured at a given inspection point
19 should be preserved. Applying a threshold comparison and
preserving only the binary result represents a loss of information
relative to cell structure characterization. Preserving magnitude
information provides greater variation in detector response, and
therefore greater utility in subsequent pattern recognition
analysis. Accordingly, detector 23 output T represents a magnitude
of light intensity measured at inspection point 19 for each
inspection point 18 scanned by the apparatus 20.
The height of surface 16, in particular inspection point 18,
can be measured by scanning apparatus 20 through the use of
commonly available triangulation techniques. As may be
appreciated, the height of inspection point 18, e.g., especially
when coincident to a crack or split in article 12, affects other
measurements taken by scanning apparatus 20. Collecting a height
datum for each inspection point 18 augments the collection of
multi-dimensional information associated with inspection point 18,
and thereby further enhances pattern recognition algorithms applied
thereto. Accordingly, a triangulation light detector 35 calculates
a height of inspection point 18 and provides its output H to the
discrimination circuit 28. Light detector 35 represents a variety
of well known methods for detecting the height of inspection point
18. For example, a detector placed at a given angle relative to
the light beam 27, e.g., 45 degrees, views inspection point 13 by
focusing against a CCD array. The location of the resulting image
19
CA 02547304 1995-04-12
on the CCD array relative to an expected location, e.g.,
corresponding to a reference height for inspection point 18,
provides representation of the height of inspection point 18. It
will be understood, however, that output H provided to
discrimination circuit 28 may be provided by a variety of well
known height detection methods.
As will be apparent to those skilled in the art, by suitably
indexing the position of article 12 relative to apparatus 20 and
collecting detector outputs S, R, T, and H in association with a
given inspection point 18, a diverse set of widely varying
information representing grain structure at the inspection point 18
is stored in a scan vector. Thus, for each indexed position of
article 12 a scan vector of mufti-dimensional data is associated
with a corresponding inspection point 18. Indexing rollers 70
contact the upper surface 16 and lower surface of article 12 and
roller control 72 moves article 12, by way of rollers 70, in
indexed fashion while providing inspection point 18 location output
X,Y to discrimination circuit 28. Discrimination circuit 28 then
associates a physical location on surface 16 with the inspection
point 18 for each indexed position of article 12. By such
association, a scan vector holding a collection of grain structure
representative data taken relative to inspection point 18 may be
associated .with a specific physical location, expressed in X and Y
dimensions, relative to wood article 12. Multiple lonqitudinal
scanning passes, but across different width portions of surface 16,
CA 02547304 1995-04-12
as indicated by lateral indexing 73, provides scan vectors for the
entire surface 16. Discrimination circuit 28 thereby constructs a
data representation 74 of surface 16 as a collection of scan
vectors.
FIG. 3 illustrates the structure of a scan vector. Each scan
vector represents a mufti-dimensional characterization of the
corresponding inspection point 18 including dimensions X, Y, S, R,
T, and I~. As may be appreciated, data representation 74 comprises
a plurality of such scan vectors, each scan vector corresponding to
one inspection point 18 of wood article 12. Discrimination circuit
28 also calculates additional fields for each scan vector as a
function of such detector outputs. For example, a calculated S/R
ratio may be incorporated into each scan vector.
In addition to data representing cell structure, traditional
image-based information is collected for each inspection point 18.
Accordingly, a white light source 46 illuminates surface 16 at the
inspection point 18. A three-color light detector 54 views
ZO inspection point 18 and, with appropriate filtering, measures light
intensity in each of a first, second, and third color band, e.g.,
measures green, red and blue light intensity. Corresponding
detector 54 outputs C1, C2, and C3 are delivered to discriminator
circuit 28 to augment characterization of the inspection point 18.
?S Accordingly, each scan vector (FIG. 3) developed by discrimination
circuit 28 in association with each inspection point 18 further
21
CA 02547304 1995-04-12
includes fields C1, C2, and C3 characterizing the color intensity
of inspection point 18 in each of a first, second, and third color
band. Use of detector 54 to detect light intensity in each of
three separate color bands is well known in the art and available
in commercial products. Generally, appropriate filtering and
processing of light intensity information, taking into account the
presence of laser light illumination at inspection point 18,
provides accurate representation of the color intensity of
inspection point 18 in each of three separate color bands.
FIG. 4 illustrates graphically the data representation 74 of
wood article 12. In FIG. 4, data representation 74 appears as a
collection of two dimensional data structures, individually labeled
74a-74h. Each of structures 74a-74h include cells addressed in X
and Y dimensions corresponding to X and Y dimensions for location
of inspection points 18 relative to wood article 12 surface 16.
Each of structures 74a-74h correspond to one of the apparatus 20
values S, R, T, H, S/R, C1, C2, and C3, respectively. Thus, a
single scan record vector resides throughout the data structures
74a-74h at corresponding X and Y locations therein. By collecting
from data representation 74 a scan vector, i . a . , a set of values at
corresponding X and Y locations in each of structures 74a-74h, one
obtains a scan vector corresponding to one inspection point 18. It
will be appreciated by those skilled in the art that a variety of
data representations may be employed to store and access a
collection of scan vectors whereby each scan vector characterizes
22
CA 02547304 1995-04-12
an inspection point 18 according to the values S, R, T, H, S/R, C1,
C2, and C3.
FIG. 5 illustrates by flow chart a first use of the data
representation 74 far a given wood article 12. In FIG. 5,
processing begins in block 200 where apparatus 20 collects scan
vectors for a given wood article 12, i.e., builds a data
representation 74. Continuing to block 202, the method of grain
defect scanning is applied to the collection of scan vectors to
locate and measure the size of grain defects. As noted above,
reflective grain defect scanning is well suited for locating and
measuring the size of grain defects such as knots. In block 204,
the result of grain defect scanning analysis is reported, i.e. , the
location and size of knots as found under grain defect scanning is
collected for later use. As may be appreciated, only a portion of
the scan vector fields need be used to execute reflective grain
defect scanning, i.e., fields S, R, and S/R of the scan vector
provide sufficient basis to execute reflective grain defect
scanning.
Continuing to block 206, those scan vectors associated with
defects found under reflective grain defect scanning are deleted
from the collection of scan vectors, i.e., removed or masked from
data representation 74. Because the corresponding portions of wood
article 12 have been accurately located and measured in size,
further consideration of this data is unnecessary. In block 208,
23
CA 02547304 1995-04-12
the method of tracheid effect analysis is applied to the remaining
scan vectors. As noted herein above, by first eliminating from
consideration scan vectors associated with grain defects located
and measured under reflective grain defect scanning, use of a
relatively less complex, i.e., lower resolution, detector 23 is
feasible. Eliminating scan vectors associated with defects found
under reflective grain defect scanning relieves the subsequent
process of tracheid effect analysis by limiting responsibility of
such analysis to identification of only broad-area defects, e.g.,
stains, wood decay, and compression wood. Such broad-area defects
are adequately identified through tracheid effect analysis using
the above-described relatively lower resolution form of detector
23. Continuing to block 210, those defects found under tracheid
effect analysis are reported.
Thus, the process illustrated in FIG. 5 uses certain data for
identifying those defects best represented by that data, and uses
other data for other defects best represented by such other data.
As may be appreciated, the process illustrated in FIG. 5 may be
?0 further modified to incorporate additional data, e.g., color
outputs C1, C2, and C3, for detection of conditions best
represented by such data.
The data representation 74 is subjected to pattern recognition
'5 analysis whereby scan vectors taken from representation 74 are
analyzed relative to redefined training set vectors established
24
CA 02547304 1995-04-12
during a preliminary training session. More particularly, a
relationship found to minimize distance among values in a training
set is applied to the values of a scan vector; and if similar
minimized distance among values in the scan vector results then the
scan vector "statistically matches" the training set vector. In
this manner, the multi-dimensional data of a scan vector, other
than the positional X and Y location data, establishes through
mathematic analysis characterization of a given inspection point
18. -
FIG. 6 illustrates the structure of a training set vector as
contemplated under the present invention. The training set vector
corresponds in structure to that of the scan vector, with the
exception of the X and Y location data and the inclusion of
additional fields representing selected mathematical functions. As
may be appreciated, the X and Y location data becomes relevant when
necessary to reference a particular location on wood article 12 in
light of a characterization thereof following pattern recognition
analysis. Each training set vector also includes a name field N
representing the wood grain condition characterized by that
training set vector. For example, a training set vector may
represent a "live knot" and the name field N identifies that
training set vector as one representing that particular wood grain
structure.' Once a training set vector is statistically matched
with a scan vector, the X and Y location data taken from the scan
vector is associated with the name field N taken from the
CA 02547304 1995-04-12
statistically matching training set vector, thereby providing
characterization of a particular physical location of article 12
surface 16.
To develop a collection of training set vectors for use in
pattern recognition applied to the data representation 74, a
"training" session is executed wherein an operator of apparatus 20
associates selected inspection points 18 with known wood grain
structures, e.g., particular types of defects or normal grain
patterns. Thus, a control station 80 (shown only in FIG. 2) drives
roller control 72 and lateral indexing 73 enabling the operator
thereof to feed sample articles 12, with examples of known defects
or clear wood conditions, through scanning apparatus 20.
Once a data representation 74 for a sample wood article is
generated by apparatus 20, the operator executing the training set
session observes the representation 74 on terminal screen 80b to
identify within data representation 74 scan vectors associated with
known defects of clear wood conditions exhibited by the sample
?0 article 12. Once such scan vectors are identified in the data
representation 74, the operator actuates a selection button 80d to
define the selected scan vectors as training set vectors for the
particular defect or clear wood condition and also enters by way of
keyboard 80.a a name designation for the selected defect or clear
'.5 wood condition. The selected training set vectors are then
reported by station 80 to discrimination circuit 28. This process
26
CA 02547304 1995-04-12
continues until the operator has defined and reported to
discrimination circuit 28 a training set vector for each defect or
clear wood condition exemplified by sample wood article 12.
In this manner, discrimination circuit 28 builds the name
field and data fields for training set vectors as represented in
FIG. 5, i.e., including a name field N designating the particular
defect or condition represented by the remaining data fields S, R,
T, H, S/R, C1, C2, and C3. As may be appreciated, multiple
training set vectors may be associated with a single name field,
i.e., the name field need not be uniaue among a group of training
set vectors. As explained more fully hereafter, additional
function fields in each training set vector identify one or more
mathematical functions found to minimize the distance between the
values held in the data fields. Function fields F1, F2, and F3
represent such one or more mathematical functions associated with
each training set vector. It will be understood, however, that one
or any number of such functions may be found to sufficiently
minimize the distance between data values held in the fields S, R,
T, H, S/R, C1, C2, and C3. Upon termination of the training
session, discrimination circuit 28 holds a collection of training
set vectors as the data representation 76.
In FIG. 7, once a training session is complete and a suitable
?5 collection of training set vectors organized as the data
representation 76, a pattern recognition process 100 receives the
27
CA 02547304 1995-04-12
data representation 74 for a particular wood article 12. By
analyzing under multivariate analysis the data representation 74,
a collection of scan vectors, against the data representation 76,
a collection of training set vectors, each inspection point 13 is
characterized. Generally, each portion of the wood article 12,
e.g., as represented by a scan vector, is analyzed by reference to
the collection of training set vectors. By employing pattern
recognition analysis 100, the scan vectors of data representation
74 are statistically matched to the training set vectors of data
representation 76. Upon finding a statistical match therebetween,
the name field N of a statistically matching training set vector is
associated with the X and Y dime..~.sicns of the matching scan vector
whereby a particular location cn a wood article 12 is associated
with a wood condition characterization, i.e., associated with the
name field N of a statistically similar training set vector. Thus,
pattern recognition analysis 100 provides as output 102 a series of
datum pairs =04 each pair comprising an X,Y location taken from a
scan vector and an associated wood conditicn characterization as
taken from a statistically matching training set vector. Given a
sequence of such datum pairs 104 for a given wood article 12, the
entire article 12 is represented by characterization of specific
physical locations thereon according to, for example, known defect
types. Subsequent processing steps, taking into account
characterization of specific locations of wood article 12
advantageously make use of or grade wood article 12 accordingly.
2a
CA 02547304 1995-04-12
71208-89
The following discussion describes generally the process of
multivariate analysis under pattern recognition analysis 100 of the
present invention. It will be understood, however, that this class
of mathematical analysis is well known and may be implemented in a
variety of ways once the scan vectors and training set vectors are
collected and prepared for use. U.S. Patent No. 5,311,131 filed
May 15, 1992 by applicant Justin P. Smith, entitled MAGNETIC
RESONANCE IMAGING USING PATTERN RECOGNITION, and issued May 10,
1994 discusses generally the use of similar pattern recognition
analysis, but applied to distinct data and in a distinct field of
use. The disclosure of U.S. Patent No. 5,311,131 may be referenced
for the processing of training set vectors and scan vectors. In
other words, the method of pattern recognition analysis proposed
under the present invention may proceed in similar fashion to that
shown in U.S. Patent No. 5,311, 131. The method shown in U.S.
Patent No. 5,311,131 has as an objective, however, the presentation
of an image whereas under the present invention system output
comprises reporting of wood conditions at specific locations on a
wood article.
Generally, each training set vector is completed by
identifying one of more functions which, when applied to the data
therein, minimizes the distance therebetween. The process of
identifying and developing such functions for representation in
fields F1, F2, and F3 is well known to those familiar with
29
CA 02547304 1995-04-12
multivariate pattern recognition analysis. To analyze a scan
vector relative to a training set vector, therefore, the data
fields of a scan vector are applied to a function taken from a
training set vector, as represented by one of fields F1, F2, and
F3, and if the result shows sufficiently minimum distance among
data points then the scan vector statistically matches the training
set vector.
Defect scanning such as described in U.S Patent 5,252,836,
when combined with motion of a scanned article in a direction
perpendicular to the scan, naturally gathers a two-dimensional
array of data from one surface of the article. This array of data
establishes multiple dimensions of data, for example the specular
reflection, retroflection, thickness, color, or tracheid effect
optical signals. At each inspection point, then, there is an n-
dimensional vector of data representing the measured values of each
of the signals from the inspection point.
one can more easily visualize data values and their
significance in classification of each inspection point 18 when the
data is one-dimensional or two-dimensional. For example, the
technique of image processing using thresholds often includes a
binary map of the inspection points 18 corresponding to the value
of one dimensional data, e.g., each pixel represents intensity of
a particular optical signal above or below a given threshold. This
procedure often corresponds to common sense visual observations
CA 02547304 1995-04-12
about the material being inspected. An example would be searching
for undesirable dark spots in a cloth fiber.
Unfortunately, when applied to lumber defect inspection, the
relative brightness of a given inspection point may have no meaning
without consideration of a substantial number of neighboring points
for their individual brightness or for the shape or size of a group
of these points that may constitute a particular defect. This
makes image processing procedures as applied to lumber defect
scanning time consuming, complex, and susceptible to error.
The present invention, however, classifies individual
inspection points 18, on the basis of the data from that point only
or from average values of the data from nearby neighbors according
to a branch of pattern recognition called multivariate analysis.
Such classifications are successful, even when the relationships or
correlations among the data ar a so complex and mul ti-dimensional to
make impossible visualization by a human being. In other words,
interrelationships among the data may be impossible to visualize,
understand, or explain, but the process of multivariate pattern
recognition correlates scan vector data and training set data to
establish characterization of an inspection point 18.
Multivariate analysis of the optical data proceeds as
described above by "training" the analysis procedure on a sample of
each of the lumber defects of interest. This training procedure,
31
CA 02547304 1995-04-12
also known as supervised classification, requires a human operator
to identify a particular region of interest, i.e., define a
training set vector. The training set vector or vectors correspond
to one or more examples of the defect to be classified. The
training set vector is selected, for example, by moving joystick
3oc to locate a cursor on computer display terminal 80b displaying
one of the optical data sets as a gray scale image.
A statistical procedure is then used to find a combination of
the multidimensional data that optimally describes the data vectors
within the training set. A number of statistical procedures are
available for this step, for example K Nearest Neighbor (KNN) or
Soft Independent Modeling by Class Analogy (SIMCA). Selection of
an appropriate procedure for a given classification problem like
lumber defect inspection involves minimizing the error in
classification by choosing the data dimensions (measured variables)
that provide the most information, and by selecting a
classification that is best suited for these data.
Classification then proceeds by applying a similar analysis to
each of the inspection points, i.e., to each scan vector (a test
set) , also selected by an operator or by considering all of the
scan vectors taken from an article 12. The result of this analysis
is a measure of distance (in n-dimensional space) of the scan
vector data points (test set points) from the training set vector
for each of the defects. This resulting "distance" (or inversely,
32
CA 02547304 1995-04-12
similarity) can be used to classify areas on surface 16 as being
one of the known defects of interest or clear wood (which can also
be identified via a training set vector). A statistical
probability for this classification can also be derived.
Pattern recognition analysis may be conducted, for example,
according to what is referred to as "cl uster analysis. " Under such
cluster analysis, data values in a training set vector define data
points in an n-dimensional space. The axes of this n-dimensional
l0 space are manipulated, i.e., rotated, until a measure of separation
between data points of the training set vector is minimized. A
variety of methods may be employed to determine distance between
data points in n-dimensional space, e.g., widest separation,
average separation, or sum of separations. Once one or several
dimensional axes rotation functions are found to sufficiently
minimize the distance between data points for a given training set
vector, these functions are associated with that training set
vector for use against data values held in a scan vector. To apply
data values of a scan vector to an axes rotation function
associated with a given training set vector, the same dimension
axes rotated by the function associated with the training set
vector are selected for rotation relative to the scan vector data
values. If rotation of the scan vector axes of rotation results in
sufficiently minimal clustering of the data points thereof, then
?5 the scan vector is considered to be similar to the training set
vector. The extent of similarity between a given training set
33
CA 02547304 1995-04-12
vector and a given scan vector may be expressed as a probability
based on the difference in clustering provided by a given axes
rotation function against the training set vector data values
versus the scan vector data values. In other words, distance is
considered the inverse of similarity.
Image processing techniques can then be applied, if necessary,
to assist in classification of these defect areas by their shape
and size. Note that these techniques are more likely to be correct
if other classification means have already been used to identify a
region or interest.
It can be appreciated that multivariate signal processing
techniques can take substantial processing time if not applied in
I5 an optimal way. Preprocessing of the data can reduce this
processing time and decrease the classification error. For
example, identifying knots using grain defect scanning techniques
and applying multivariate analysis or other classification
techniques like image processing to a limited set of defects (not
including knots) for the remaining surface of the board under
inspection improves accuracy and processing time.
It will be appreciated that the present invention is not
restricted 'to the particular embodiment that has been described and
illustrated, and that variations may be made therein without
departing from the scope of the invention as found in the appended
34
CA 02547304 1995-04-12
claims and equivalents thereof. For example, U.S. Patent No.
5,252,836 shows a more practical form of reflective grain defect
scanning provided by a flying-spot laser scanner. It will be
understood, therefore, that additional detectors shown herein,
e.g., providing outputs H, T, C1, C2, and C3, may be incorporated
into a flying-spot laser scanner in the manner in which detectors
22 and 24 are shown in U.S. Patent No. 5,252,836 in the form of a
flying-spot laser scanner.
