Computers and Electronics in Agriculture 118 (2015) 204–214

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

An automated method for tree-ring delineation based on active contours

guided by DT-CWT complex coefficients in photographic images:
Application to Abies alba wood slice images
Pol Kennel a,b,⇑, Philippe Borianne a, Gérard Subsol b
a
Agricultural Research for Development, UMR AMAP, F-34398 Montpellier, France
b
Laboratory of Informatics, Robotics and Microelectronics, ICAR Team, CNRS/University Montpellier, F-34000 Montpellier, France

a r t i c l e i n f o a b s t r a c t

Article history: This paper describes an efficient method for delineating tree-rings and inter tree-rings in wood slice
Received 13 March 2015 images. The method is based on an active contour approach and a multi-scale gradient map resulting
Received in revised form 6 September 2015 from the Dual Tree Complex Wavelet Transform (DT-CWT). The method is automated and does not
Accepted 8 September 2015
require any pith localization. It is also quite robust to some defect structures such as branch prints, cracks,
Available online 21 September 2015
knots or mold. We applied the method to process entire Abies alba wood slices (aged from 10 to 50 years)
from bark to pith, which amounted to about 200 tree-rings. Our automatic delineation method performed
Keywords:
accurately compared to the manual expert measurements with a mean F-score of 0.91 for the quality of
Tree-ring delineation
Active contour
delineation.
Dual-tree complex wavelet transform Ó 2015 Elsevier B.V. All rights reserved.
Dendrochronology

1. Introduction quality, especially strength and stiffness (Hanning et al., 2003).

Tree-ring surfaces can be used for geometrical measurements,
In temperate trees, wood production by the cambium is rhyth- which can be related to mechanical properties of the wood. For
mic, with a new layer of wood being added each year. This iterative example, asymmetric growth will result in non-circular and irreg-
wood production process generates a set of circular shapes on the ularly spaced tree rings, which will characterize what is called
cross-sectional plane of a stem which are called ‘‘tree-rings”. Tree- reaction wood, which is a combination of compression and tension
ring boundaries are visible thanks to a difference of wood density wood. Characterizing and quantifying wood tissues paves the way
in the ring. More specifically, late wood on the outer portion of the for 3-dimensional models of wood growth, which is an important
ring is darker and denser than the early wood formed on the inner issue for understanding biomass accumulation and carbon seques-
portion of the ring. The study of tree-rings is a common and funda- tration. All these parameters could help industrialists to predict
mental part of dendrology and helps to understand, quantify and whether a log is suitable for plywood, paper or plank making
model secondary tree growth (MacDonald, 1998). (Kumpulainen and Marjanen, 2010).
In most cases, one-dimensional radial intensity profiles (from Some computer-aided methods have been proposed to count
bark to pith) of photographic images of a wood slice are used to tree-rings and to give an estimation of their thickness in photo-
detect tree-ring boundaries. This allows one to calculate the tree graphic images of wood slices. They can detect some parts of rings
age and to obtain valuable information on its environmental in a portion of an image, as in (Conner et al., 1998), by using an
conditions such as climate or forest dynamics (MacDonald, edge detector or detect intensity discontinuities along a ray
1998). Nevertheless, if we want to study the growth process of defined from the pith to the bark, as in (Vaz et al., 2004; Norell,
the tree in greater depth, we need precisely to delineate the entire 2011), or the commercial software WinDendro (Regent Instru-
tree-ring areas in wood slices. Thus, the width and density (in ments, Canada1). While those techniques provide valuable parame-
numbers) of tree-rings show how biomass is allocated in the ters for dendrochronology applications, they do not lead to accurate
woody material, resulting in valuable information about wood segmentation and quantification of whole tree-rings. Other algo-
rithms have been proposed to analyze the shape of the tree-ring
structure in order to detect internal defects. For example, in
⇑ Corresponding author at: Agricultural Research for Development, UMR AMAP,
F-34398 Montpellier, France.
1
E-mail address: [email protected] (P. Kennel). http://www.regentinstruments.com/assets/windendro_about.html.

http://dx.doi.org/10.1016/j.compag.2015.09.009
0168-1699/Ó 2015 Elsevier B.V. All rights reserved.
 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214 205

(Bhandarkar et al., 2005) the authors proposed characterizing cracks cannot deal with changes in topology. Implicit formulation, intro-
by detecting lines making a significant non-zero angle with the rings. duced in (Caselles et al., 1993; Malladi et al., 1995), solved this
The method, based on an edge detector and line fitting procedure, problem. It consists, as in the level set formulation (Suri et al.,
provides efficient detection of the ring structure. Nevertheless, the 2002; Giraldi et al., 2006), in embedding the active contour as
rings themselves are not precisely delineated and, in particular, they the zero level set of a higher dimensional function and in solving
are not closed. the corresponding motion equation. The main advantage of
In fact, few automatic techniques have been developed to delin- implicit representation is its ability to change the contour’s
eate complete tree-rings precisely. Most of them first require topology during deformation. The main advantage of parametric
detection of the pith, which is a key problem. For this purpose, in representation is its effectiveness. For example, vertex sampling
(Sjöberg et al., 2001) the authors proposed correlating a vector may not be constant or uniform on parametric contours, whereas
image (which more or less represents a local orientation) with an resolution on implicit contours is constrained by the resolution of
tecnicas para detectar
image characterizing an accumulation of concentric circles. The a regular grid. Implicit active contour models generally require
centros
method described in (Chalifour et al., 2001) computes the gradient considerable computational power, and parametric active contour
in a large number of randomly chosen pixels and approximates the models are difficult to set up and control. Simpler models based
pixel toward which all the gradient orientations point. In (Tadeusz on a discrete set of points were therefore introduced. In (Gao
et al., 2003), a center pixel is selected which enables conversion of et al., 1998), the point-based active contour consists of an
the image into polar coordinates. When the center pixel corre- ordered collection of points, called ‘‘snaxels”, renowned for their
sponds to the pith, the polar image is formed by horizontal lines simple implementation and usage. In the study described here,
which can be easily characterized. In (Cerda et al., 2007), the center we opted for the point-based approach proposed in (Delingette
is defined as the pixel which minimizes an objective function and and Montagnat, 2001), where a framework is defined for
which fits all the points belonging to any detected edge to concen- handling the active contour shape and topology in a simple and
tric circles. In (Borianne et al., 2010), the Hough transform is used intuitive manner. The real issue is then to determine an efficient
to detect circles by using many diameter values. Once the pith is external force.
found, radial directions can be defined to enhance the intensity External forces are established in such a manner as to minimize
variations around the ring pixels as in (Borianne et al., 2010) or the ‘‘energy” of the active contour. For image segmentation, the
deteccion de
detect edges or intensity discontinuities as in (Chalifour et al.,
anillos idea is to have low energy values around object boundaries. The
2001; Sjöberg et al., 2001) or (Cerda et al., 2007). It is worth noting basic approach uses the image gradient as an external force to
that in (Soille and Misson, 2001), there is no assumption on pith attract the active contour to the object boundary (Kass et al.,
localization. The method is only based on latewood to earlywood 1988). More sophisticated approaches offer successive refinements
transitions, which are detected using the Watershed algorithm. from coarse to fine segmentation (Nagappan et al., 2008). Several
Another difficulty is to link the pixels that are detected as belong- authors have offered the possibility of coupling these refinements
ing to a tree-ring, in order to obtain a closed curve. Mathematical with wavelet transforms (Nagappan et al., 2008; Shan and Ma,
Linkeo de
morphology operators puntos
such as erosion, dilation and skeletonization 2010). For example, the model in (Dah Way et al., 2005; Zhang
make it possiblepertenecientes a anillos
to fill holes and to produce one pixel width lines, and Liu, 2005) reduces computation time thanks to the multi-
as in (Sjöberg et al., 2001; Soille and Misson, 2001). In (Cerda et al., scale analysis nature of the wavelet transforms. In this case,
2007), the authors propose working directly with a set of polygons because of their different singularities, signals and noise are easily
which are concentrically included and whose vertices are pixels distinguished by the local maxima of their coefficients. Wavelet
belonging to the rings. In (Borianne et al., 2010), it is a closed line Vector Flow (WVF) was then compared with a classic Gradient
that is directly fitted to some pixels tagged as ‘‘ring” points by Vector Flow (GVF) and showed to be effective on few medical
using the active contour paradigm. This kind of method, which is images segmentation (Cheng et al., 2007). Therefore, as wavelet
widely used in image processing, was first described in a paper transform has been widely and successfully used to capture
by (Kass et al., 1988), in which we can already find an application multi-scale edges in medical images (Wang and Yang, 2007), we
to wood slice photographs. Using a closed line directly solves the propose to evaluate its performance in wood slice images for the
problem of linking isolated pixels and increases the robustness of purpose of tree-ring delineation. We opted for a relatively recent
the method with respect to local wood defects. enhancement of the Discrete Wavelet Transform (DWT) called
Our study once again addressed the question of using active the Q-Shift Dual Tree Complex Wavelet transform (DT-CWT)
contours to delineate concentric tree-rings. because of its near shift invariance and its frequency analysis qual-
Active contour models (also called ‘‘snakes”) are often used to ities (Kingsbury, 2001). DT-CWT has already been successfully
bridge the gap between low-level computer vision and high-level used in the level set approach (De Rivaz and Kingsbury, 2000)
geometric representation. An active contour is a curve described where it represented surface areas by complex coefficients derived
within an image domain which can move under the effect of from the transformation. In our snake approach, we introduce a
internal forces regularizing the shape of the curve, and external new external force based on a feature derived from the DT-CWT:
forces derived from the image attracting the curve toward some the Inter Coefficient Product (ICP) extracted from the DT-CWT
regions of interest. An active contour is often presented as an which provides efficient features suitable for an active contour
energy minimization model, which integrates a number of con- strategy by yielding edge orientations in the filter support regions
straints derived from specific forces. Active contour models have (Anderson et al., 2005). We then summarize the different scales
been widely used in medical image processing (Cai et al., 2006; and orientation sub-band features provided by the transform into
Nagappan et al., 2008), texture segmentation (Chen et al., 2006) a single gradient-map that can be used as an attractor-map to
or visual tracking (Paragios and Deriche, 2000). Active contour design the external force.
methods are quite difficult to classify, as shown in (McInerney Our contribution is therefore as follows:
and Terzopoulos, 1996; He et al., 2008), but we can define three
major categories: parametric, implicit and point-based models. In
We propose a new approach for coupling the Q-Shift Dual Tree
the original parametric formulation (Kass et al., 1988), the Complex Wavelet Transform and the point-based active contour
optimal contour position was defined as that with the lowest model, and more particularly propose a formulation of the
sum of internal and external energies along the entire length of external force based on the magnitudes and orientations given
the contour. In this formulation, the continuous active contour by the ICP.
206 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214

We introduce different notions, such as attraction power and

attraction orientation, which can be used to determine the most
appropriate attractors and thereby simulate an external force
field that results in rapid convergence to the balanced position.

We describe and test a new method to delineate successive
concentric tree-rings in Abies alba flatbed-scanned images by
running successive active contour delineating tree-rings and
inter tree-rings alternately.

2. Description of the method

2.1. The Dual-Tree Complex Wavelet Transform (DT-CWT)

Fig. 1. DT-CWT impulse response filters: row (a) presents the real parts and row (b)
In matters of texture image analysis and synthesis, the Q-Shift the imaginary parts of complex filters, row (c) presents the magnitude of the filters.
Dual-Tree Complex Wavelet Transform (DT-CWT) (Kingsbury,
2001), originating from the wavelet community, has proven to be
superior to all other wavelet transforms (De Rivaz, 2000; Lo
et al., 2011). The DT-CWT proposed by Kingsbury in (Kingsbury, orientation map which we then consider as a wavelet-based
2001) is a relatively recent enhancement of the Discrete Wavelet edge-map. The strategy provides a correct edge characterization
Transform (DWT) where the DWT is considered to be one of the of the image studied by merging the multi-scale edge analysis
most useful mathematical tools for multi-resolution image result from the DT-CWT. The proposed procedure is as follows:
analysis (Mallat, 1989). DWT aims to break down (decompose) a
signal into a decimated multi-scale representation that isolates 1. Decompose the image to w by the N-level DT-CWT.
coarse image components into a sparse set of coefficients. It has 2. Compute ICP coefficients for each oriented sub-band of the n-
been used in a wide range of applications for handling object levels.
detection, and image de-noising, compression and segmentation 3. Scale each sub-band to the size of the original image by interpo-
(Lu et al., 1997; Kim and Kang, 2007). The DT-CWT comes with sev- lation. Note that this step is not necessary when using the
eral desirable properties for feature extraction: these transforms undecimated version of the DT-CWT (Hill et al., 2012).
have similar shapes to Gabor wavelets (De Rivaz, 2000), with 4. At each level, summarize the six oriented sub-bands to wn (with
important properties additional to the DWT. It exhibits approxi- n the scale level) by computing, at each location (x, y), the max-
mate shift invariance, is directionally selective (p/10, p/4, 2p/5, imum value of ICP magnitudes over the six sub-bands and by
3p/5, 3p/4, 9p/10 vs 0, p/2 and 3p/4 for the DWT) and achieves keeping the corresponding phase value.
perfect reconstruction along with good frequency localization. 5. Summarize the N sub-bands obtained in step 4 to the final edge
The main advantages of the DT-CWT over Gabor wavelets and map wf by computing, at each location (x, y), the vector sum of
other separable filter banks are limited redundancy and greater each complex coefficient by:
computational efficiency. !
X
N
The 1-dimensional DT-CWT is implemented by using a pair of argðwf ðx; yÞÞ ¼ arg ðjwn ðx; yÞj  argðwn ðx; yÞÞÞ
filter banks that process the input signal simultaneously. One tree n

produces the real and the other the imaginary part of the complex
coefficient. For the 2-dimensional DT-CWT, which is similar to a 2- X
N

dimensional DWT, the tree pair is applied to the rows then the col- jwf ðx; yÞj ¼ jwn ðx; yÞj
n
umns of the input image. This operation results in six complex
high-pass sub-sampled sub-bands and one real low-pass sub-
band at each level. Ultimately, the size of the final sub-bands is The orientations and magnitudes found at each level of the
1/2k the size of the original signal for a k-level DT-CWT decompo- decomposition (step 4) are summarized in step 5 with the aim of
sition. Fig. 1a presents the impulse responses of the six complex using the most relevant orientation in the final features map. Thus,
wavelets associated with 2-D analysis filters. the proposed vector product considers the N orientations weighted
by their corresponding magnitudes. This results in a complex final
2.2. Defining features using Inter Coefficient Products (ICP) from the map where the phase presents an optimal orientation map from
DT-CWT the multi-scale analysis, and where the magnitude can be used
as an attractor feature in subsequent sections. Fig. 3 illustrates
A relatively simple and computationally efficient representation these two maps considered for the remainder of the study on an
of the coefficient phases of the DT-CWT coefficients provided Abies alba wood slice image sample; color images are converted
access to the angles of dominant directional features in their into gray levels by averaging color canals before the gradient com-
support regions. These consist of the so-called Inter Coefficient putation. Such a gradient map cannot be used directly for tree-ring
Products (ICP) (Anderson et al., 2005) which are computed by mea- delineation by thresholding high gradient values. Discontinuities
suring the relative phase of appropriate neighbor coefficients for and local contrast variations do not allow a correct segmentation
the oriented sub-bands considered. These ICP coefficients are then so that an active contour process is used to fit the tree-ring gradi-
used to construct Texture Orientation Maps (TOM) as illustrated in ent continuously.
Fig. 2. This new feature is used in (Vo and Oraintara, 2010) for
texture classification and segmentation by characterizing ICP 2.3. Integrating the features in the active contour framework
phase distribution.
In this work, we want to evaluate the potential benefits of For the active contour framework, we chose a point-based rep-
ICP features as an external force for an active contour model. resentation which was proposed by Delingette and Montagnat
For this purpose, we summarize ICP features into a single texture (2001). At a time t, the closed contour is defined by a list of n
 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214 207

Fig. 2. ICP coefficients from the DT-CWT of image (a): (b) and (c) represent the magnitude and the phase of the complex coefficients at level 3 of the decomposition in the
sub-band oriented at 2p/5. In (d), local features are plotted as an arrow oriented according to phase, and are elongated according to magnitude.

Fig. 3. Wood slice sample photographic image (a), with the magnitude map (b) (the histogram is stretched for visualization), and the orientation map (c) (gray levels
represent angles comprised between p/2 and p/2) obtained from the DT-CWT transform and the summary of ICP coefficients.

connected points P ti such that P tn is linked to P t1 . The evolution of In Eq. (2), the function L( ) is given by the following expression:
the active contour is then described by the equation:
 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
    ri
Ptþ1 ¼ Pti þ ð1  cÞ Pti  Pt1
t
þ ai Finti þ bi Fextti ð1Þ L r i ; ui ; ei ¼ 1 þ l 1 þ 4ei ð1  ei Þtan2 ui
i i tanui

where ai and bi are respectively the weights for the internal and the where
external forces applied to point Pti and c is a general damping term
which allows oscillations to be avoided. In order to simplify the l ¼ 1 if jui j < p=2 and l ¼ 1 if jui j > p=2
notation, in the remainder of this paper we write Pi (respectively L is the normal component of the internal force and fixes the elas-
t
Finti , Fext i ) instead of Pti (respectively Finti , Fextti ) for the vertex ticity of the snake. When consecutive points of the snake are
position at time t. aligned, ui is nil and the L term is arbitrarily set to zero.
The internal function Fint i , regularizes the positions of the The external force Fexti in Eq. (1), at point P i , can be defined as
points with respect to the others through the following equation: an attraction force toward a point called the attractor and called Ai.
  Fexti is defined as energy due to the interaction of the snake with
Finti ¼ i Pi1 þ ð1  i ÞP iþ1 þ L ri ; ui ; i ni ð2Þ
the image. The definition of Ai is closely related to the application.
where In our approach, we used the DT-CWT-based gradient described in

ei and (1  ei) are the barycentric coordinates of the projection the previous section. Ai is based on the magnitude and orientation
of Pi onto the line [Pi1, Pi+1]. maps given by the DT-CWT. The aim is to find attractors, i.e. suffi-

ri is the half distance between the two neighbors Pi1 and Pi+1 of ciently strong local maxima on the magnitude map, along the
Pi. directions locally given by the orientation map. Once Ai is estab-

ui⁄ is the oriented simplex angle which gives the continuity con- lished (see next section), the (Pi Ai) couple forms a normalized vec-
straint. It is defined as follows tor which is weighted by the unit coefficient determined from a
mean combination between the distance and intensity of the
X
j¼iþs
uj attractor. The external force decreases as point Pi converges toward
ui ¼ its balanced position.
j¼is
ð2s þ 1Þ
In the point-based active contour, the external function Fexti is
uj being the oriented angle between the two segments [Pj1, Pj] as follows:
and [Pj, Pj+1], and s the coefficient of rigidity. It defines the  
1 jdi j
smoothing of the snake. 8i 2 f1; . . . ; ngFexti ¼ Dv i þ wi
2 dmax

ni is the local unit contour normal vector, and is oriented toward
the exterior. where
208 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214

Dvi is the unit value difference, or magnitude difference the wood slice border and let it converge toward the first tree-
(derived from the DT-CWT-based gradient) between Pi and its ring. Then, iteratively (1) we changed the external force by taking
attractor Ai. It should be noted that these magnitudes are values the minimal values in the gradient-map which corresponded to the
scaled between 0 and 1. di is the signed distance between Pi and so-called inter tree-ring position and we let the active contour con-
its attractor Ai: di is positive when Ai is outside the contour and verge toward this area (see Fig. 4, in blue), (2) we set to zero all
negative when it is inside. dmax is a parameter that regulates the magnitudes of the gradient-map which were external to the cur-
distance maximum beyond which attraction is considered to be rent position of the active contour which was now stabilized
nil. around a tree-ring, (3) we used the external force defined in Sec-

wi is the unit attraction vector; it is conventionally defined in tion 2.3 in order that the active contour could move toward the
the literature as the normal vector on the Pi position oriented next (internal) tree-ring (see Fig. 4, in red). The process allowed
to the exterior of the contour. Here, we define it as the weight- us to delineate tree-rings from the bark to the pith for an entire
ing vector sum wood slice image as shown in Fig. 5 on a 48-year-old Abies alba
wood slice. The method takes a reasonable average time of around
wi ¼ ð1  v i Þni þ v i oi
5 min. on a standard machine depending mostly on the image res-
oi is the unit vector given by the orientation map. It gives the olution, the number of tree-rings, the number of contour’s points
direction of the closest contour detected by the wavelets. When and the image complexity (in term of structure and intensity). Note
the vi magnitude at the Pi position is nil or too weak, the attrac- that the tree-ring search was stopped when the pith was encoun-
tion occurs along the normal vector ni to the snake; when the vi tered, i.e. when the inner distribution variance of the contour
magnitude is well tagged, the attraction occurs along the orien- reached a given threshold. More robust constraints could be
tation given by the wavelets. This vector sum simulates a force envisaged either on the contour (radius, perimeter, surface, etc.)
field without any extrapolation or diffusion of the values given or on its image/gradient domain (statistics, texture, etc.). As well,
by the DT-CWT orientation map. an automated initialization of the active contour around the first
tree-ring could be envisaged for example by background and bark
Finally, Fext is designed such that the greater the potential dif- detection.
ference is between Pi and its attractor Ai, the stronger the attraction
is, and such that the greater the distance is between Pi and its
3. Results
attractor Ai, the weaker the attraction is.
The potential Ai points along the attraction axis are defined as
3.1. Application to Abies alba wood slice images
the most attractive local maxima computed in a mask of a given
length k on both sides of the contour, and allow both internal
When applied to tree-ring delineation, our method yielded
and external displacements.
accurate results as seen in Fig. 5. In particular, the active contour
An attraction power is calculated for each potential attractor Ai
adopted the expected position for tree-ring delineation even when
found, and is related to the current position Pi. The final selected
some wood defects were present. Fig. 6 shows the results around
attractor is that which has the greatest attraction. Attraction power
branch insertions (Fig. 6a and d), knots (Fig. 6c) and small and
pa is a signed coefficient that expresses the following idea: attrac-
medium cracks resulting from log drying (Fig. 6d and b,
tion decreases as the distance between the point and its attractor
respectively).
increases; attraction increases as the intensity difference between
the point and its attractor increases. The attraction power function
thus depends on the positions and magnitudes of compared points 3.2. Evaluation metric
(by convention, the attraction power of two superimposed points
is nil). This definition is used to choose the most appropriate In order to assess the method with respect to contours delin-
attractor when several local maxima are found on both sides of eated by experts, to analyze the influence of some parameters or
the contour. Attraction power is defined by the following to quantify sensitivity to image noise, we had to compare different
expression: pixel contours for a given tree-ring. But, we could not just study
vi  vj the overlapping of contours as, for example, two co-circular con-
8 Pi – P j paij ¼ tours shifted by only one pixel may not have shared any pixel. This
kPi P j k
demonstrates the need to introduce a margin error to consider that
The result of our method is a polygonal line [P 0 , P i , Pn ] which two or more contours are identical (or at least consistent). For
represents a tree-ring delimitation. This polygonal line can be example, pixels can be put in correspondences between contours
transformed into a closed 8-connected relationship contour in via a min-cost assignment method on a bipartite graph and given
the image. a margin error (Scott and Nowak, 2006). However, we opted for a
global geometrical approach to avoid the pixel matching procedure
2.4. Application to tree-ring delineation which is complex. Our comparison between two contours was
based on the overlapping of regions which were created by thick-
The proposed method was used to detect tree-rings in images of ening each contour. For this purpose, we applied to each contour a
Abies alba wood slices. Tree-rings were delineated from the bark to mathematical morphology operator called dilation, which consists
the pith. We have seen that a tree-ring corresponds to a sudden in aggregating pixels in the neighborhood of the existing pixels. If
transition between a light and a dark area and that its boundary we performed n dilations, the contour had a thickness of 2n + 1
corresponds to the highest values in the gradient-map. The idea pixels which defined the region corresponding to a margin error
was therefore to start the active contour in a light area, i.e. corre- of n.
sponding to the lowest magnitude values in the gradient-map, Hence we were able to define Precision/Recall measurements.
located around a tree-ring, in order that the external force defined First, we defined gt and s as, respectively, the contour which was
in Section 2.3 could be effective. Of course, to avoid any ambiguity, considered as ground truth (in general, the tree-ring delineated
it was essential to remove from the gradient map any external by an expert) and the one given by our method. We were then able
magnitudes which could also attract the active contour. For this to compute, gt d ðnÞ and sd ðnÞ; which were the same contours after n
purpose, we first initialized manually the active contour along dilations. Then, we defined True Positives as TP ¼ s \ gt d , Falses
 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214 209

TP, FP, TN and FN depend on the margin error n. A large margin

error provides good detection scores, which may be not significant
for experts, whereas too small a margin error always results in
poor detection scores. Thus, n has to be calibrated with great care.
For this purpose, three different experts delineated manually the
same 5 rings of an Abies alba image by using the same functional-
ities of the same software, ImageJ (Rasband, 1997–2014) and by
following the same protocol (e.g. zoom level during annotation,
approximate space between points of the polygonal line, etc.).
The idea was that a significant margin error to assess our method
should be in the range of the average inter-expert error. As we did
not know the exact contour of a tree ring, we compared all the
pairs of contours given by the experts. For one tree-ring (and a
given n), we therefore had 6 F-scores that were averaged. We pre-
sent in Fig. 7a the evolution of the mean F-score in line with the
margin error n (in pixels). If we assumed that a mean F-score of
0.9 showed a good agreement between the 3 experts, we found
Fig. 4. Successive delineations of tree-rings (in red) and inter tree-rings (in blue) that we had to select a margin error of n = 6 pixels, which corre-
using our method. Tree-rings (respectively inter tree-rings) are given by the active sponded to 0.4 mm in the image (see an illustration of this margin
contours which are stabilized on local maxima (respectively local minima) in the in the case of delineation by two experts in Fig. 7b). We used this
displayed gradient-map. (For interpretation of the references to colour in this figure
value for the assessment of all the following experiments since our
legend, the reader is referred to the web version of this article.)
images had similar resolutions.

Positives as FP ¼ s n TP, True Negatives as TN ¼ sd \ gt d and False 3.3. Evaluation of the DT-CWT-based gradient
Negatives as FN ¼ sCd \ gt. Finally, the Precision measurement was
given by P ¼ TP=ðTP þ FPÞ, the Recall measurement by We chose a priori the DT-CWT for (1) its multi-resolution ability
R ¼ TP=ðTP þ FNÞ and the F-score by F ¼ 2PR=ðP þ RÞ. In the follow- to provide relevant magnitudes in a wide range of tree-rings
ing, we use the F-score, which provides a balanced average of the combined with (2) computing ICP which are efficient in providing
Precision and Recall ranged between 0 (worst score) and 1 (best accurate edge (or tree-ring transition) orientations on extensive
score), to characterize the quality of tree-ring delineations. areas, (3) while being robust to noise. In this section, we address

Fig. 5. Full delineation of tree-rings on an image of a 46-year-old Abies alba wood slice. The parameters of the snake were set as: number of points = 300, a ¼ 0:5, b ¼ 0:5,
k = 50.

Fig. 6. Our method succeeded in delineating accurate tree-rings even in the presence of defects such as branch prints (a, d), cracks (b, d) and knots (c).
210 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214

Fig. 7. (a) Mean F-score resulting from the delineations made by 3 different experts on the same 5 tree-rings. The margin error corresponds to the number of dilations n
applied on the discrete contours in order to compute the TP, TN, FP and FN coefficients; (b) margin error (with n = 6) on two expert delineations, in yellow and green
respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

the relevance of the DT-CWT-based gradient with respect to more concentric in (Fig. 3). Indeed, we obtained a near constant
standard features. For this purpose, we applied the proposed DT- DT-CWT gradient orientation, which was expected for an accurate
CWT procedure, as well as the classic Sobel operator, on the wood guidance of the active contour.
slice presented in Fig. 3. First, we took a 1-dimensional radial pro- Hence, the Sobel gradient is very sensitive to noise; if it is to be
file from the bark to the pith and we present in Fig. 8 the original used to define an external force for the active contour, a blurring
intensity profile (upper row), the Sobel gradient magnitude (mid- filter needs to be applied on its magnitude map. This reduces the
dle row) and the DT-CWT-based gradient magnitude (bottom influence of the noise but at the cost of some loss of precision in
row). In addition, both gradient orientations can be seen in light the maxima coordinates. Without a blurring step, the Sobel
gray. It can be seen that the DT-CWT gradient is much smoother gradient delivers high magnitudes that are too imprecise and
than the Sobel gradient, avoiding high magnitude values resulting misplaced so that it is not possible to process the entire wood slice
from local noise. Furthermore, in a radial cross-section, tree-ring from bark to pith with successive delineations as described in
transitions are expected to be oriented similarly as they are quite Section 3.1.

Fig. 8. Results on a1-dimensional radial profile from bark to pith extracted from the wood slice presented in Fig. 3. Upper row: original intensity profile. Middle and bottom
rows: Sobel gradient and DT-CWT-based gradient profiles, magnitudes are in black and orientations in gray. The DT-CWT-based gradient offers smoother values and delivers
a near constant gradient orientation. In particular, around x = 130 there is a very deep and narrow valley in the original intensity profile which characterizes a punctual noise.
At this value, the Sobel magnitude is very high whereas the DT-CWT magnitude value remains stable, showing the robustness of the DT-CWT transform with respect to noise.
 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214 211

Fig. 9. Evolution of F-score with respect to the number of iterations for several snakes with the Sobel-based gradient (a) and DT-CWT-based gradient (b). In (c), the
corresponding mean F-score evolution for both gradients (Sobel with continuous red line and DT-CWT with continuous blue line); in addition, the percentage of stabilized
snakes for the Sobel-based gradient (with dashed redline) and DT-CWT-based gradient (with dashed blue line). (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)

We then wanted to compare the accuracy of both gradients. F-score evolution obtained for both gradients (in blue DT-CWT, in
Thus, we processed an entire wood slice and compared 28 tree- red Sobel) as well as the proportion of stabilized active contours
ring delineations based on the DT-CWT gradient and on the Sobel (out of the 28): i.e. contours which reached their final position with
gradient. In both cases, we used strictly the same parameters for no more moving points. First, it shows how the DT-CWT-based gra-
the active contour: the number of points was 300, a ¼ 0:5 and dient outperformed the Sobel-based gradient in terms of delin-
b ¼ 0:5. The Sobel magnitudes map was blurred with a Gaussian eation precision. Furthermore, it can be seen that the active
kernel (d ¼ 12) so that the successive tree-ring delineation process contour stabilized much quicker with the DT-CWT gradient than
worked well. Our results showed a mean F-score of 0.88 ± 0.06 with the Sobel-based gradient, e.g. at iteration 700, half of the active
with the DT-CWT gradient and a mean F-score of 0.66 ± 0.22 with contours were stabilized with the DT-CWT-based gradient while
the Sobel gradient. none were with Sobel-based gradient. This resulted from the robust
We present in Fig. 9 the F-scores of the 28 active contours at dif- and relevant directions that were given by the phase of the DT-
ferent steps (iterations) of their evolution until convergence. Both CWT-based gradient, as well as smooth magnitude values.
gradients, Fig. 9a and b for the Sobel and DT-CWT gradients respec- Finally, if we look at the Sobel-based active contour behaviors,
tively, are plotted. Additionally, we summarize in Fig. 9c the mean we can see that most of them went through on relevant positions

Fig. 10. (a) F-score distribution resulting from the automatic tree-ring delineation over the full set of seven Abies alba wood slices, for a total of 204 tree-ring delineations.
(Min = 0.52, 1st Quartile = 0.89, Median = 0.94, Mean = 0.91, 3rd Quartile = 0.96, Max = 1.00). (b) Illustration of delineations scored low (F-score equal to 0.54 and 0.57) and (c)
scored high (F-score equal to 0.87 and 0.92). Expert delineations are in green (b, c), automated delineations are in red. The illustrations show a transposed version of the
original images after a polar transform in order to facilitate visual estimation. (For interpretation of the references to colour in this figure legend, the reader is referred to the
web version of this article.)
212 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214

Fig. 11. Results of our method applied on six images of Abies alba wood slices of different ages presenting specific default structures such as branch prints, cracks, knots or
mold.

(with high F-scores) but did not stabilize on them and continued on (because of blurring or due to residual noise). On the other hand,
to wrong positions (with weaker F-scores). This was mainly due to DT-CWT-based delineations stabilized well on optimal localiza-
high local maxima which were misplaced on the gradient-map tions (see Fig. 10).
 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214 213

3.4. Comparison with expert measurements simply tuned by the number of points and the stop condition is
determined by the number of moving points and/or the maximal
To assess the suitability of our method, we asked experts number of iterations. Vertex spacing is controlled through the tan-
(i.e. botanists specialized in wood anatomy) to manually annotate gential component of the internal force applied to each vertex and
7 images with different tree ages (from 10 to 50 years) and with we opted for an equidistant spatial distribution in order to con-
several defects highlighted previously (cracks, mold prints, branch strain the distance between successive vertices. Internal force
prints, non-circular tree-ring, etc.). We then obtained 204 ground- expression does not depend on contour parameterization: it regu-
truth tree-rings. Fig. 11 illustrates our image set (except the one larizes the contour curvature without producing shrinkage. Such a
presented in Fig. 5) as well as the delineations resulting from the point-based active contour could also manage topology changes:
proposed method. In Fig. 10a, we can see the distribution of the (Delingette and Montagnat, 2001) proposed an algorithm that cre-
F-score between the ground-truth contours and the contours given ates and merges different connected components of the contour
by our method. The accuracy yielded by our automatic method was based on the detection of edge intersections. In particular, this
comparable to that of the expert measurements with a mean should be of particular usefulness for processing wood slices with
F-score of 0.91 ± 0.08, considering a margin error of 6 pixels. These several piths. Moreover, as no lengthy global minimization calcula-
results highlight the accuracy of the method compared to the tions are necessary to determine an optimal ‘‘stabilized” position,
inter-expert error measured using the same margin-error (see Sec- the convergence process remains fast.
tion 3.2); in most cases, our method delineated tree-rings with the We developed a method that can deal with Abies alba wood slice
same accuracy as an expert could have achieved. Fig. 10b,c illus- images. But the complexity of the content in such images varies
trate samples of the automated delineation with the expert’s delin- and mainly depends on the type of wood analyzed and the image
eation for two cases where F-scores reached by our automatic acquisition system. Particular wood cell types may be more pro-
method were 0.55 and 0.90. nounced (e.g. rays and vessels) and can disturb the tree-ring delin-
eation method. An important issue in this context is also to deal
with branch insertions, off-centered second pith, false tree-rings,
4. Discussion and conclusion knots and confused tree-rings. In the last case, the successive
tree-ring and inter tree-ring detection process described here
This study provides a full framework for automatically delineat- could run into problems due to irregular annual (or periodic) wood
ing tree-rings in photographic images, which may solve major growth. Ways of improvement may reside in the a priori detection
problems in dendrology, including, among other things, the detec- of these particular wood cells and structures, depending on the
tion of structures such as knots (Longuetaud et al., 2012) and piths species, in order either to adapt gradient computation or the active
(Boukadida et al., 2012). The point-based active contour method contour evolution scheme.
chosen, coupled with a DT-CWT-based gradient, proved its
capacity to delineate tree-rings in images of Abies alba wood slices. Note
We assessed the precision of the automatic delineation method by
providing encouraging Precision Recall measurements yielded We provide, as supplementary materials, our set of images as
between several automatically-delineated and expert-delineated well as expert delineations and our delineation results. The pro-
tree-rings. posed method will be available in the TOASTER2 platform.
In most cases, contour initialization conditions are an important
issue. Our method yielded good results by starting with normal
Acknowledgments
orientations and appropriate ranges for maxima. However,
improvements could be made by using a continuous vector flow
We wish to thank O. Taugourdeau for providing the Abies alba
approach that could be provided by the DT-CWT through a conven-
wood slice images (wood slices were prepared and photographed
tional diffusion process (Xu and Prince, 1998a,b; Cheng et al.,
by P. Cretin-Maitenaz and the INRA-URFM technical team) and
2007). This was not necessary in our particular context but it could
M. Jones for assistance with the preparation of the manuscript.
be relevant in other fields. The comparison itself between General-
ized GVF (Xu and Prince, 1998b) and our approach may be of inter-
est in a more general context since we showed that ICP features Appendix A. Supplementary material
can provide local vector flow. In our method, segmentation accu-
racy mainly depended on the gradient provided by the DT-CWT Supplementary data associated with this article can be found, in
analysis. Multi-scale edge detection is another major issue in the the online version, at http://dx.doi.org/10.1016/j.compag.2015.09.
field of computer vision. In our study, DT-CWT clearly demon- 009.
strated its ability to capture relevant contour localization on mul-
tiple scales and orientations. The merging scale solution proposed References
here is simple and efficient compared to the other methods we
considered, such as reconstructing the image after setting low- Anderson, R., Kingsbury, N., et al., 2005. Determining multiscale image feature
angles from complex wavelet phases. Comput. Sci. 3656, 490–498.
pass coefficients to 0, or reconstructing the image while holding
Bhandarkar, S.M., Luo, X., et al., 2005. Detection of cracks in computer tomography
only local maxima in high-pass coefficients (Mallat and Zhong, images of logs. Pattern Recogn. Lett. 26 (14), 2282–2294.
1992). However, the number of scales to be considered for the Borianne, P., Pernaudat, R., et al., 2010. Automated delineation of tree-rings in X-Ray
transform is an open question. We systematically used 6 levels of computed tomography images of wood. In: 2011 18th IEEE International
Conference on Image Processing (ICIP).
decomposition in this study but the precision of local maxima Boukadida, H., Longuetaud, F., et al., 2012. PithExtract: a robust algorithm for pith
localization is directly affected by this parameter. Other wavelet detection in computer tomography images of wood – application to 125 logs
transforms could be considered in this context, particularly the from 17 tree species. Comput. Electron. Agric. 85, 90–98.
Cai, H., Xu, X., et al., 2006. Repulsive force based snake model to segment and track
contourlet and shearlet transforms (Wang and Yang, 2007), which neuronal axons in 3D microscopy image stacks. NeuroImage 32 (4), 1608–1620.
deal efficiently with edge localization. Caselles, V., Catté, F., et al., 1993. A geometric model for active contours in image
The point-based active contour method employed in this study processing. Numer. Math. 66 (1), 1–31.
has several advantages in the delineation of tree-rings. In particu-
lar, it is simple to parameterize. For example, contour rigidity is 2
http://amap-collaboratif.cirad.fr/pages_logiciels/index.php?page=toaster.
214 P. Kennel et al. / Computers and Electronics in Agriculture 118 (2015) 204–214

Cerda, M., Hitschfeld-Kahler, N., et al., 2007. Robust tree-ring detection. Advances in Lu, C.S., Chung, P.C., et al., 1997. Unsupervised texture segmentation via wavelet
Image and Video Technology. Second Pacific Rim Symposium, PSIVT 2007, transform. Pattern Recogn. 30 (5), 729–742.
Santiago, Chile, December 17–19. Springer. MacDonald, G.M., 1998. Tree rings and environment: dendroecology. By F.H.
Chalifour, A., Nouboud, F., Deprost, B., Okana S., 2001. Automatic detection of tree- Schweingruber. New Phytol. 138 (4), 743–750.
rings on wood slice images. In: 5th International Conference on Quality Control Malladi, R., Sethian, J.A., et al., 1995. Shape modeling with front propagation: a level
by Artificial Vision, Le Creusot France. set approach. IEEE Trans. Pattern Anal. Mach. Intell. 17 (2), 158–175.
Chen, S., Sochen, N.A., et al., 2006. Integrated active contours for texture Mallat, S., Zhong, S., 1992. Characterization of signals from multiscale edges. IEEE
segmentation. IEEE Trans. Image Process. 15 (6), 1633–1646. Trans. Pattern Anal. Mach. Intell. 14 (7), 710–732.
Cheng, J., Liu, Y., et al., 2007. A new active contour model for medical image analysis Mallat, S.G., 1989. A theory for multiresolution signal decomposition: the wavelet
– wavelet vector flow. Int. J. Appl. Math. 36. representation. IEEE Trans. Pattern Anal. Mach. Intell. 11 (7), 674–693.
Conner, W.S., Schowengerdt, R.A., et al., 1998. Design of a computer vision based McInerney, T., Terzopoulos, D., 1996. Deformable models in medical image analysis:
tree ring dating system. In: IEEE Southwest Symposium on, Image Analysis and a survey. Med. Image Anal. 1 (2), 91–108. http://dx.doi.org/10.1016/S1361-
Interpretation, 1998. 8415(96)80007-7, ISSN: 1361–8415.
Dah Way, F., Rajeswari, M. et al., 2005. Segmentation using wavelet and GVF snake. Nagappan, A., Prabhu Britto, A., et al., 2008. Novel segmentation technique using
ACIT – Signal and Image Processing. Novosibirsk, Russia. wavelet based active contour model for detection of mammographic lesions.
De Rivaz, P., 2000. Complex Wavelet Based Image Analysis and Synthesis. Inf. Technol. J. 7, 490–496.
University of Cambridge. Norell, K., 2011. Automatic counting of annual rings on Pinus sylvestris end faces in
De Rivaz, P., Kingsbury, N., 2000. Fast segmentation using level set curves of sawmill industry. Comput. Electron. Agric. 75 (2), 231–237.
complex wavelet surfaces. In: Proceedings, 2000 International Conference on, Paragios, N., Deriche, R., 2000. Geodesic active contours and level sets for the
Image Processing, 2000. detection and tracking of moving objects. IEEE Trans. Pattern Anal. Mach. Intell.
Delingette, H., Montagnat, J., 2001. Shape and topology constraints on parametric 22 (3), 266–280.
active contours. Comput. Vis. Image Underst. 83 (2), 140–171. Rasband, W.S., 1997–2014. ImageJ. USA, U.S. National Institutes of Health.
Gao, J., Kosaka, A., et al., 1998. A deformable model for human organ extraction. In: Scott, C., Nowak, R., 2006. Robust contour matching via the order-preserving
Proceedings, 1998 International Conference on, Image Processing, 1998, ICIP 98. assignment problem. IEEE Trans. Image Process. 15 (7), 1831–1838.
Giraldi, G., Suri, J., et al., 2006. Implicit dual snakes for medical imaging. Conf. Proc. Shan, H., Ma, J., 2010. Curvelet-based geodesic snakes for image segmentation with
IEEE Eng. Med. Biol. Soc. 1 (1), 3025–3028. multiple objects. Pattern Recogn. Lett. 31 (5), 355–360.
Hanning, T., Kickingereder, R., et al., 2003. Determining the average annual ring Sjöberg, P.-J., Danielsson, P.-E., et al., 2001. Image analysis of annual ring pattern for
width on the front side of lumber. prediction of wood quality. In: Symposium on Image Analysis, Norrpöking,
He, L., Peng, Z., et al., 2008. A comparative study of deformable contour methods on Sweden.
medical image segmentation. Image Vis. Comput. 26 (2), 141–163. Soille, P., Misson, L., 2001. Tree ring area measurements using morphological image
Hill, P., Achim, A., et al., 2012. The Undecimated Dual Tree Complex Wavelet analysis. Can. J. For. Res. 31 (6), 1074–1083.
Transform and its application to bivariate image denoising using a Cauchy Suri, J.S., Kecheng, L., et al., 2002. Shape recovery algorithms using level sets in 2-D/
model. In: 2012 19th IEEE International Conference on Image Processing (ICIP), 3-D medical imagery: a state-of-the-art review. IEEE Trans. Inf. Technol.
Orlando, USA. Biomed. 6 (1), 8–28.
Kass, M., Witkin, A., et al., 1988. Snakes: active contour models. Int. J. Comput. Tadeusz, S., Morel, O., et al., 2003. Detection automatique des stries de croissance
Vision 1 (4), 321–331. des arbres par transformÃÓe en ondelettes. CNRIUT.
Kim, S.C., Kang, T.J., 2007. Texture classification and segmentation using Vaz, C., Carvalho, P., et al., 2004. A vision-based system for automatic growing ring
wavelet packet frame and gaussian mixture model. Pattern Recogn. 40 (4), detection and measurement. Comput. Ind. Eng. 46 (2), 347–354.
1207–1221. Vo, A., Oraintara, S., 2010. A study of relative phase in complex wavelet domain:
Kingsbury, N.G., 2001. Complex wavelets for shift invariant analysis and filtering of property, statistics and applications in texture image retrieval and
signals. Appl. Comput. Harmonic Anal. 10 (3), 234–253. segmentation. Signal Process.: Image Commun. 25 (1), 28–46.
Kumpulainen, P., Marjanen, K., 2010. Directional variance analysis of annual rings. J. Wang, P.S.P., Yang, J., 2007. A review of wavelet-based edge detection methods. Int.
Phys: Conf. Ser. 238 (1), 012047. J. Pattern Recognit Artif Intell. 26 (07), 1255011.
Lo, E.H.S., Pickering, M.R., et al., 2011. Image segmentation from scale and rotation Xu, C., Prince, J.L., 1998a. Snakes, shapes, and gradient vector flow. IEEE Trans.
invariant texture features from the double dyadic dual-tree complex wavelet Image Process. 7 (3), 359–369.
transform. Image Vis. Comput. 29 (1), 15–28. Xu, C., Prince, J.L., 1998b. Generalized gradient vector flow external forces for active
Longuetaud, F., Mothe, F., et al., 2012. Automatic knot detection and measurements contours. Signal Process. 71 (2), 131–139.
from X-ray CT images of wood: a review and validation of an improved Zhang, H.-W., Liu, Z.-G., 2005. Wavelet-based snake model for image segmentation.
algorithm on softwood samples. Comput. Electron. Agric. 85, 77–89. In: Proc. of SPIE 6044, pp. 604420-604420.