Annals of Botany 78 : 325–331, 1996
Automatic Classification of Plant Cells According to Tissue Type using
Anatomical Features Obtained by the Distance Transform
A. J. T R A V I S, D. J. H I R S T* and A. C H E S S O N
Rowett Research Institute and *BioSS, Rowett Research Institute, Greenburn Road, Bucksburn,
Aberdeen AB2 9SB, UK
Received : 11 December 1995
Accepted : 8 March 1996
Image analysis was used to develop a faster and more objective method for the quantitative measurement of plant
anatomy. The size, number and anatomical features of individual cells were measured automatically using methods
based on image skeletonization and the distance transform. The variation in cell wall thickness around individual cells
was measured by masking the distance transform of a segmented binary cell wall image with a skeleton of the same
image to extract distance values along the boundary between neighbouring cells. The measurements were used to
create a multi-dimensional feature space in which individual cells were classified automatically according to their
tissue type. Cells were classified into seven types using discriminant analysis, and the performance of the classification
rule was examined by cross-validation. The potential use of these methods as a research tool, and in plant breeding
programmes is discussed.
# 1996 Annals of Botany Company
Key words : Image analysis, plant anatomy, wood anatomy, cell classification, epidermis, sclerenchyma, xylem,
parenchyma, vascular tissue, maize.
INTRODUCTION
The distribution of cell wall material between different plant
tissues has a direct influence on the performance and
suitability of a particular crop species or variety for specific
end uses. However, routine, quantitative measurement of
plant anatomy has proved too time-consuming and laborious to find widespread application where a large number
of samples are involved as, for example, in a plant breeding
programme. If anatomical traits are to be considered in
plant breeding or in the assessment of crop quality, faster
and more objective methods are needed to quantify
anatomical features. Delineation of different tissues within
a thin section of plant material is normally based on the
presence of groups of cells with anatomical features in
common. The size and shape of cells, and the thickness and
uniformity of their walls are the main distinguishing
characteristics. Automatic delineation of tissues may be
possible on the basis of reliable classification of individual
cells according to tissue type. This has become increasingly
practical as the cost of image analysis equipment has fallen.
Inexpensive systems capable of advanced feature extraction
are now readily available (Moss, 1988), and can be used to
measure the anatomical features of individual plant cells.
Although, at present, many of these systems require
significant operator intervention to make anatomical
measurements on individual cells, quantitation of wood
anatomy is used routinely as a method of assessing the
quality of wood as timber, or as a source of fibre for
pulping. Ilic and Hillis (1983) developed a low-cost image
analysis system for quantification of features such as lumen
0305-7364}96}090325­07 $18.00}0
area, and the proportional area occupied by various cell
types of Eucalyptus regnans which was tedious and difficult
using traditional methods. Similar methods were used by
Gasson (1985) to measure the anatomical features of xylem
vessels in Oak (Quercus robur) and Beech (Fagus sylŠatica).
Image analysis has also been used as an alternative to X-ray
densitometry in assessing latewood density at the cellular
level in Pinus ponderosa (Park and Telewski, 1993).
Methods of image analysis used for quantitation of wood
anatomy are now increasingly directed towards measurement of the variation in anatomical features of individual
fibres (Evans et al., 1993 ; Evans, 1994) which is an important
factor in predicting the paper-making qualities of wood
pulp. In particular, the work of Peachey and Osborne (1990)
on the quantitation of wood anatomy established the
‘ distance ’ transform (Borgefors, 1986) as a useful method
of measuring the variation in anatomical features of
individual tracheids, and of automatically excluding ray
parenchyma.
The importance of anatomical features in determining the
properties of crop plants is similarly recognized. Forage
anatomy, for example, plays a major part in determining the
accessibility of cell walls to rumen micro-organisms and
thus the rate at which ingested plant material is degraded
(Wilson, 1993 ; Wilson and Mertens, 1995). Although most
anatomical studies of crop plants are essentially qualitative
in their approach, image analysis methods based on the
distance transform have been used to measure the anatomical features of individual maize cells (Travis, Hirst and
Chesson, 1994 a), loss of cell wall material from ‘ thick ’
(50 µm) sections of maize internode during in Šitro digestion
# 1996 Annals of Botany Company
326
TraŠis et al.—Automatic Classification of Plant Cells
by cell wall degrading enzymes (Travis et al. 1994 b), and
anatomical features in Gerbera jamesonii stems (Van der
Heijden, Van de Vooren and Weil, 1995).
Measurement of the extent of natural variation in
anatomical features of existing crop varieties, and the
identification of traits suitable for genetic manipulation are
important in establishing the range of anatomical phenotypes that could be produced. The objective of the work
presented here is to develop a system capable of classifying
large numbers of cells automatically on the basis of their
anatomical features, with minimum operator intervention,
for eventual use in the selection of improved crop varieties
and in the evaluation of existing or alternative sources of
plant fibre for industrial use.
MATERIALS AND METHODS
Microscopy
Measurements of anatomical features were made in five
different areas of a transverse section of Zea mays internode
(Public cell line W401, silage stage) stained with Toluidine
Blue, at a magnification of ¬400 using a Zeiss Axioskop
microscope (Zeiss, Germany). The section was fixed in 10 %
glutaraldehyde in 0±1  phosphate buffer, dehydrated in a
series of increasing concentrations of ethanol then embedded
in LKB Historesin (Leica, Milton-Keynes, UK) and
sectioned (5 µm) as described by Travis, Murison and
Chesson (1993).
Image analysis
A Hitachi KP-104 monochrome CCD camera (Hitachi
Japan) fitted to the microscope was used to obtain digital
video images of the tissue section. The images were digitized
at a resolution of 512¬512 pixels (picture elements) and 256
grey-levels using a Sun Sparcstation 10 (Sun Microsystems,
UK) and a Prima Graphics Virtuoso frame-grabber (Prima
Graphics, Royston, UK). The anatomical features of
individual cells in the digitized images were measured using
the ‘ Visilog 4 ’ image analysis system (Noesis, France).
A digitized image of the internode section (Fig. 1 A) was
first segmented into cell walls and background using an
automatic thresholding algorithm (Riddler and Calvard,
1978). The binary cell wall image produced (Fig. 1 B) was
then spatially filtered using ‘ morphological ’ operators to
remove high-frequency noise (Serra, 1982). A morphological
‘ opening ’ (erosion followed by dilation) of one pixel was
used to remove small foreground objects and irregularities
in the larger object boundaries, then a morphological
‘ closure ’ (dilation followed by erosion) of one pixel was
used to remove small holes from the remaining foreground
objects (cell walls).
A skeleton of the binary cell wall image which approximates to the position of the middle lamella between
adjacent cell walls (Travis et al., 1993) was obtained using
the ‘ watershed ’ algorithm (Fig. 1 C). The ‘ watershed ’
algorithm identifies local maxima in the distance transform
of the binary cell wall image (Fig. 1 F). The binary
‘ watershed ’ was then reduced to a single-pixel width 8connected skeleton (i.e. neighbouring pixels in all eight
directions belong to the same object). The resulting skeleton
is equivalent to obtaining the pruned skeleton of the
background.
The discrete geometry of a digitized image on a square
sampling grid produces a paradox (Hilditch, 1969). By
convention, foreground objects are regarded as 8-connected
and the background as 4-connected. The connectivity of the
skeleton was therefore changed from 8 to 4 using a
morphological thickening in two orthogonal directions. The
4-connected skeleton produced was then inverted and used
to segment the image into discrete 8-connected foreground
‘ regions ’ corresponding to individual cells in the original
grey-level image. Morphological operators were used to fill
holes in the foreground regions, and any edge-connected
components of the image were removed to ensure that
measurements were only made on cells completely within
the field of view.
Feature extraction
The 8-connected regions (cells) were labelled (Fig. 1 D)
and their features measured individually. The area, perimeter, shape (perimeter#}4π area), Euler number, Feret
diameters, and mean, variance, range, maximum and
minimum grey-level of individual cells were measured on
the overall, cell lumen, cell wall and distance images. In the
case of distance images, the ‘ grey ’ value corresponds to the
mean, variance or range of cell wall thickness for each cell.
on the initial ‘ overall ’ label image using the Visilog ‘ analysis ’
module (Noesis, 1993). The ‘ overall ’ label image (Fig. 1 D)
was then ‘ masked ’ using the binary cell wall image (Fig. 1 B)
to produce a ‘ cell wall ’ label image (Fig. 1 E). The ‘ cell wall ’
label image and the original grey-level image (Fig. 1 A) were
used to obtain the area and grey-level features of the cell
walls. A second shape descriptor and related measures were
also obtained from this image. The ‘ overall ’ label image was
masked by the inverse binary cell wall image to produce a
‘ lumen ’ label image which was used to measure the area and
grey-level features of the cell lumen. The ‘ overall ’ label
image (Fig. 1 D) was ‘ masked ’ with a binary image of the
cell perimeters and used in conjunction with the distance
transform of the binary cell wall image (Fig. 1 F) to obtain
estimates of mean cell wall thickness and the variation in cell
wall thickness of individual cells.
Distance transform
The ‘ grey ’ value in the distance image (Fig. 1 F)
corresponds to the radius of a circle from each foreground
pixel to the nearest background pixel, and was obtained
using morphological operators to erode the foreground
objects (cell walls). The number of erosions required to
remove a pixel is the distance value for that pixel. A single
line of the original grey-level image can be represented by a
one-dimensional signal (Fig. 2). The dashed line represents
the one-dimensional equivalent of the boundary detected in
a single line of the grey-level image (Fig. 2 A) by an
automatic threshold algorithm to produce a binary signal
(Fig. 2 B). The ‘ distance ’ transform of this binary signal
TraŠis et al.—Automatic Classification of Plant Cells
327
F. 1. A, Grey-level image of a transverse section of Zea mays internode (5 µm, ¬1000 oil immersion) stained with Toluidine Blue. B, Greylevel image segmented automatically into cell walls (shown in green) and background then filtered using morphological operators to remove highfrequency noise. C, Skeleton of binary cell wall image (shown in green) approximating to the position of the middle lamella. The skeleton was
modified, and inverted to disconnect cells from each other. D, Overall label image showing disconnected cells on which the anatomical features
were measured. E, Cell wall label image produced by ‘ masking ’ the binary image of the cell walls and used to obtain wall area, lumen area and
grey-level features. F, Distance transform of binary cell wall image. The grey value in the distance image corresponds to the radius of a circle from
each pixel in the foreground to the nearest background pixel (normalized in this image to make the distance values visible).
(Fig. 2 C) corresponds to the distance from each pixel to the
nearest boundary. The distance value reaches a maximum at
the mid-point of the cell wall. The one-dimensional skeleton
of the binary signal (Fig. 2 D) coincides with the ‘ watershed ’
of the distance transform. Using the skeleton as a mask, the
distance values along the mid-point of the cell walls were
extracted from the distance transform of the binary cell wall
image as illustrated for the one-dimensional signal in Fig.
2 E.
Statistical analysis
The results were analysed using the GENSTAT 5
statistical package (Lawes Agricultural Trust, 1987). A total
328
TraŠis et al.—Automatic Classification of Plant Cells
200
image (grey value)
A
1
segmented image (binary)
B
r
distance transform (grey value)
C
1
skeleton (binary)
D
r
E
radius (grey value)
F. 2. One-dimensional distance transform. A grey-level (0–255) signal (A) is segmented into foreground (1) and background (0) to produce a
binary signal (B). The ‘ distance ’ transform of the binary signal (C) corresponds to the radius (r) of a circle from each pixel to the nearest boundary
which reaches a maximum at the mid-point of the foreground. The skeleton of the binary signal (D) coincides with the ‘ watershed ’ of the distance
transform. Using the skeleton as a mask, the distance values along the mid-point of the cell walls can be extracted (E).
of 36 anatomical features were measured in five fields of
view. Cells were classified manually into seven ‘ true ’ groups
by visual inspection of the digitized image during measurement of the anatomical features. The cells were then
classified automatically using linear discriminant analysis of
all 36 anatomical features. Cells were classified into the
following categories : epidermis (epi), sclerenchyma (scl),
parenchyma (par), thin-walled vascular tissue (vas), xylem
vessel (xyl), air space (air) and ‘ artefact ’ (art) including
knife marks, breaks or tears in the section, and non-cell wall
material. The contribution of the anatomical features to the
derived canonical variates was expressed by weighted vector
loadings adjusted to unit standard deviation. The communalities of the feature variables (expressed as the sum of
the squared loadings for the first three canonical variates)
were used to assess the relative importance of the anatomical
features measured in discriminating between different cell
types.
RESULTS AND DISCUSSION
The mean values and coefficients of variation of all 36
anatomical features measured were examined. Although
clear differences were evident between cell types for many
features, overall cell shape and cell lumen shape were similar
in all cell types. The large variation in size of parenchyma
cells and xylem vessels combined with their low frequency of
occurrence in the section contributed to high coefficients of
variation for these tissue types. As might be expected, a
great deal of variation was also present in the features of
artefacts. The Euler number provided little information
about difference between cell types, but a large variability in
Euler number was observed in the ‘ distance ’ image for
artefacts and may be important in excluding these observations from other classes. A large variation in Euler
number for ‘ vascular ’ tissue was also observed in the ‘ cell
wall ’ image. This was a consequence of sample aliasing
between the dimensions of thin-walled vascular tissue and
the pixel resolution used.
An initial classification was made using the entire data set
to classify itself. In this case, all observations in the training
set were used to create the discriminant functions, which
were then used to classify the observations. This method has
an optimistic bias, so the classification was repeated using
cross-validation. Here each observation was classified using
discriminant functions calculated from all the data in the
training set excluding the test observation. The latent vector
loadings from the canonical variate analysis, and communalities of the feature variables for the first three
canonical variates are summarized in Table 1.
Lumen perimeter was the anatomical feature containing
most information about differences in cell type. The ‘ area ’
feature of the ‘ distance ’ (perimeter) image is an indirect
estimate of overall cell perimeter and ranked second in
importance. Cell wall perimeter was also an important
feature along with the Feret diameter of the lumen. In all, a
combination of cell and lumen size appear to be the
dominant features in discriminating between cell types.
However, other features are also of interest. In particular,
the shape of the cell perimeter in the ‘ distance ’ image was of
similar importance to the cell wall perimeter feature. The
‘ overall ’ grey level of the cell including the cell lumen and,
to some extent, the variation in grey level between cell wall
and lumen were important in discriminating cell types as
was mean cell wall thickness. Variation in cell wall thickness,
or variation in the grey-level of either cell walls or the cell
lumen contributed little to discrimination between the cell
types encountered in the training set used for this work.
The cross-validation method used is unbiased, but has a
high variance when the numbers of observations in the
groups are small. This was true in the case of xylem vessels
TraŠis et al.—Automatic Classification of Plant Cells
T     1 . Latent Šector loadings from canonical Šariate
analysis of features describing the size and shape of the cell
oŠerall, cell lumen, cell walls and Šariation in wall thickness
around the cell. The first three canonical Šariates (CV1-CV3)
are shown which explain 80 % of the Šariation. The Šariables
are adjusted to haŠe unit within-group Šariance, and the Šector
loading indicates the relatiŠe importance of each feature for
that canonical Šariate. Communalities of the feature Šariables
are expressed as the sum of squared loadings for CV1-CV3
Image
Overall
Feature
Area
Perimeter
Shape
Euler Number
Feret 0
Feret 90
Grey mean
Grey variance
Grey range
Cell lumen Area
Perimeter
Shape
Euler Number
Feret 0
Feret 90
Grey mean
Grey variance
Grey range
Cell wall Area
Perimeter
Shape
Euler Number
Feret 0
Feret 90
Grey mean
Grey variance
Grey range
Distance
Area
Perimeter
Shape
Euler Number
Feret 0
Feret 90
Grey mean
Grey variance
Grey range
CV1
(32 %)
CV2
(28 %)
CV3
(20 %) Communality
5±77
®5±14
5±56
0±00
®5±51
®3±37
®3±58
®10±30
®1±22
4±08
37±07
®1±93
0±00
®8±70
®21±45
18±86
6±64
®7±77
14±00
16±70
®1±36
®2±68
®5±51
®3±37
®10±32
0±28
5±99
13±77
®6±96
®27±92
®0±46
®5±51
®3±37
®23±81
3±83
®1±41
®5±24
®5±43
®0±55
0±00
5±79
9±16
23±09
®17±76
10±35
®7±99
®46±83
®1±89
0±00
22±87
10±37
®2±25
1±07
®5±67
12±91
®27±22
16±33
®0±65
5±79
9±16
®13±62
®3±86
2±90
®38±64
®2±33
34±52
0±57
5±80
9±16
®23±77
4±56
1±30
®2±64
®29±83
10±21
0±00
8±31
12±55
®35±10
11±34
®7±85
®5±22
95±23
®14±25
0±00
®2±44
®33±41
10±60
2±04
®1±09
13±66
35±01
®25±06
1±69
8±32
12±55
5±72
®6±89
1±93
®55±91
®29±94
®3±17
0±13
8±32
12±55
®10±63
1±67
®5±81
68
946
135
0
133
253
1778
550
170
108
12636
210
0
605
1684
473
49
94
549
2246
897
10
133
253
325
62
48
4809
950
1981
1
133
253
1245
38
37
where only seven examples were present in the training set.
There is also scope for reducing the number of redundant
variables, which may improve the performance of the
classifier. The difference between cell types is shown in
Table 2 which gives the Mahalanobis distance between the
groups (the distance standardized by the within group
standard deviation). In all cases, the group means were
separated by at least three standard deviations from the
other cell types. The separation between the groups for the
first two canonical variates is shown in Fig. 3. Further
separation of the groups is present along unseen axes
representing the other canonical variates, but the overlap
between parenchyma cells and xylem vessels is clearly
shown.
The performance of the statistical classifier was assessed
329
T     2 . Separation between groups of different cell types in
the multi-dimensional feature space (Mahalanobis distance :
the distance between the groups standardized by the withingroup standard deŠiation)
epi scl par vas xyl air art
epi
scl
par
vas
xyl
air
art
0±0
3±8
5±8
5±7
8±3
6±4
6±0
0±0
3±7
3±6
7±3
3±7
4±8
0±0
4±3
5±8
5±2
6±3
0±0
7±7 0±0
4±5 8±0 0±0
5±3 8±7 4±8 0±0
by estimating the percentage of cells correctly classified by
cross-validation. The results are shown in Table 3. The poor
performance of the classifier in recognizing xylem vessels
was surprising in view of the ease with which xylem vessels
are identified manually. The majority of the small sample
of xylem vessels measured were incorrectly classified as
parenchyma. This illustrates an important limitation of the
present approach where anatomical features alone have
been used to classify individual cells out of context from the
anatomical features of neighbouring cells. In isolation,
xylem vessels appear anatomically similar to thick-walled
parenchyma in the maize internode used for this experiment
but they are clearly identifiable as xylem vessels when seen
in their normal anatomical context. The statistical classifier
was more successful in discriminating between other cell
types. In view of the relatively small size of the training set
an average, cross-Šalidated success rate of almost 80 % for
epidermis, sclerenchyma, parenchyma and ‘ vascular ’ cell
types indicates that anatomical features containing sufficient
information to discriminate between cell types can be
measured automatically. Van der Heijden et al. (1995)
emphasized the importance of identifying and excluding airspaces from measurements of cell wall dimensions. In the
present work only seven objects out of 437 measured (1±6 %)
were incorrectly classified as air-spaces, without operator
intervention. Artefacts (knife marks, breaks, and tears in
the section) were less reliably identified.
The accuracy of the system may be improved by
incorporation of other anatomical features. Measurement
of the number of ‘ triple-points ’ (nodes in the skeleton), for
example, would allow better discrimination to be made
between xylem vessels and large parenchyma cells. In
addition, preliminary work using an artificial neural network
classifier (Travis et al., 1995) suggests that, in the longer
term, a non-linear approach might provide better discrimination than the statistical classifier used here. Stepwise
refinement of the initial classification on the basis of
anatomical context will be needed regardless of the
classification strategy adopted. Additional information
about the classification of neighbouring cells could be used
to confirm or reject an initial classification using a set of
decision rules, and a knowledge-base of the appropriate
plant anatomy.
Once reliable classification is achieved the information
provided could readily be adapted to meet the needs of a
330
TraŠis et al.—Automatic Classification of Plant Cells
6
4
CV2
2
0
–2
–4
–6
–4
–2
0
CV1
2
4
6
F. 3. Canonical variate analysis (variance explained : CV1, 32 % ; CV2, 28 %). The variables were adjusted to have unit within group standard
deviation (large circles represent a region within two standard deviations of the mean for each group of observations). (D) Epidermis ; (E)
sclerenchyma ; (^) parenchyma ; (_) vascular tissue ; (*) xylem ; (+) air space ; (X) artefact.
T     3 . Performance of the statistical classifier. Number of
cells from each true group classified statistically into seŠen
types using cross-Šalidation. Each obserŠation was classified
using discriminant functions calculated from the training set
with the test obserŠation omitted
Statistical classification
True group
epi scl par vas xyl air art Total (% correct)
epi
scl
par
vas
xyl
air
art
Total
21
2 1
13 144 9
0
7 33
0
8 4
0
0 4
0
5 0
0
2 0
34 168 51
0
8
0
65
0
0
2
75
0
0
3
0
2
0
0
5
0
12
0
5
1
66
5
89
0
2
0
2
0
2
9
15
24
188
43
84
7
73
18
437
(88)
(77)
(77)
(77)
(29)
(90)
(50)
wide range of agronomic objectives that have an anatomical
component including studies of susceptibility to lodging and
drought in cereals, and prediction of the nutritive value of
feeds and forages. Analysis of wood microstructure has
been suggested as a prospective strategy in the identification
of suitable genetic material and environments for new
plantations of trees (Evans, Gartside and Downes, 1995). A
similar strategy may be appropriate in the identification of
suitable genetic material for improved varieties of agricultural crops. Habib, Shah and Inayat (1995) suggest that
wheat varieties combining the desirable characteristics of a
high grain yield and superior quality straw can be identified
on the basis of morphological characteristics due to
genotype. Measurement of anatomical characteristics would
provide more detailed information, and may allow the
selection of genotypes to be made more accurately.
The practical demands of quality evaluation or the
routine monitoring of anatomical traits in a breeding
programme will require information about cell types to be
incorporated into a rule-based system in which individual
cells are unambiguously assigned to different tissues.
Reliable classification of cell types is a prerequisite for the
construction of the rule-based system in which the first and
most important level of decision-making is the ability to
recognize the tissue type of individual cells. The work
described here clearly demonstrates that it is possible to
produce a system capable of classifying cells automatically
on the basis of their anatomical features with minimum
operator intervention.
A C K N O W L E D G E M E N TS
This work was supported by funding from the Scottish
Office Agriculture, Environment and Fisheries Department
(SOAEFD). We thank Sandra Murison at the Rowett
Research Institute for providing the maize internode section,
and Dr Mike Franklin of BioSS at the Rowett Research
Institute for helpful discussions about the statistical aspects
of the work.
TraŠis et al.—Automatic Classification of Plant Cells
LITERATURE CITED
Borgefors G. 1986. Distance transformations in digital images.
Computer Vision, Graphics and Image Processing 34 : 344–371.
Evans R. 1994. Rapid measurement of the transverse dimensions of
tracheids in radial wood sections from Pinus radiata. Holzforschung
48 : 168–172.
Evans R, Downes G, Menz D, Stringer S. 1993. Rapid measurement of
variation in tracheid transverse dimensions in a Pinus radiata tree.
47th Appita Annual General Conference 1 : 113–119.
Evans R, Gartside G, Downes G. 1995. Present and prospective use of
Silviscan for wood microstructure analysis. 49th Appita Annual
General Conference 1 : 91–95.
Gasson P. 1985. Automatic measurement of vessel lumen area and
diameter with particular reference to pedunculate oak and common
beech. IAWA Bulletin 6 : 219–237.
Habib G, Shah SBA, Inayat K. 1995. Genetic manipulation in
morphological characteristics, chemical composition and in Šitro
digestibility of straw from different wheat cultivars. Animal Feed
Science and Technology 55 : 263–274.
Hilditch CJ. 1969. Linear skeletons from square cupboards. Machine
Intelligence 4 : 403–420.
Ilic J, Hillis WE. 1983. Video image processor for wood anatomical
quantification. Holzforschung 37 : 47–50.
Lawes Agricultural Trust. 1987. Genstat 5 reference manual. Oxford :
Oxford University Press.
Noesis. 1993. Visilog 4 reference guide. Velizy : Noeisis Vision Inc.
Moss VA. 1988. Image processing and image analysis. Proceedings of
the Royal Microscopical Society 23 : 83–88.
Park W, Telewski FW. 1993. Measuring maximum latewood density by
image analysis at the cellular level. Wood and Fiber Science 25 :
326–332.
Peachey TC, Osborne CF. 1990. The measurement of wood cell
331
parameters using the distance transform. Wood and Fiber Science
22 : 388–403.
Riddler TW, Calvard S. 1978. Picture thresholding using an iterative
selection method. IEEE Transactions on Systems, Man and
Cybernetics SMC-8 : 630–632.
Serra J. 1982. Image analysis and mathematical morphology. London :
Academic Press.
Travis AJ, Hirst D, Chesson A. 1994 a. Automatic classification of
tissues in plants of agro-industrial importance. Proceedings of the
Fifth International Congress for Computer Technology in Agriculture, Cambridge : Royal Agricultural Society of England, 157–158.
Travis AJ, Hirst D, Hamilton A, Edwards PJ, Chesson A. 1995.
Statistical and neural network approaches to the automatic
classification of cell types. Proceedings of the SeŠenth Cell Wall
Meeting, Santiago, Spain.
Travis AJ, Murison S, Chesson A. 1993. Estimation of plant cell wall
thickness and cell size by image skeletonization. Journal of
Agricultural Science, Cambridge 120 : 279–287.
Travis AJ, Murison SD, Chesson A, Perry P. 1994 b. Quantitative
assessment of cell wall degradability in forages and crop residues
using 3-D microscopy. Food Structure 1994—Abstracts, Chicago :
Scanning Microscopy International, 1–2.
Van der Heijden GWAM, Van de Vooren JG, Van de Weil CCM. 1995.
Measuring cell wall dimensions using the distance transform.
Annals of Botany 75 : 545–552.
Wilson JR. 1993. Organisation of forage plant tissues. In : Jung HG,
Buxton DR, Hatfield RD, Ralph J, eds. Forage cell wall structure
and digestibility. Madison : American Society of Agronomists,
1–32.
Wilson JR, Mertens DR. 1995. Cell wall accessibility and cell structure
limitations to microbial digestion of forage. Crop Science 35 :
251–259.