Efficient Bark Recognition in the Wild
R
´
emi Ratajczak
1,2,3
, Sarah Bertrand
1
, Carlos Crispim-Junior
1
and Laure Tougne
1
1
Univ Lyon, Lyon 2, LIRIS, F-69676 Lyon, France
2
Unit
´
e Cancer et Environnement, Centre L
´
eon B
´
erard, Lyon, France
3
Agence de l’Environnement et de la Ma
ˆ
ıtrise de l’Energie, Angers, France
Keywords:
Bark Recognition, Texture Classification, Color Quantification, Dimensionality Reduction, Data Fusion.
Abstract:
In this study, we propose to address the difficult task of bark recognition in the wild using computationally
efficient and compact feature vectors. We introduce two novel generic methods to significantly reduce the
dimensions of existing texture and color histograms with few losses in accuracy. Specifically, we propose
a straightforward yet efficient way to compute Late Statistics from texture histograms and an approach to
iteratively quantify the color space based on domain priors. We further combine the reduced histograms in a
late fusion manner to benefit from both texture and color cues. Results outperform state-of-the-art methods
by a large margin on four public datasets respectively composed of 6 bark classes (BarkTex, NewBarkTex), 11
bark classes (AFF) and 12 bark classes (Trunk12). In addition to these experiments, we propose a baseline
study on Bark-101, a new challenging dataset including manually segmented images of 101 bark classes that
we release publicly.
1 INTRODUCTION
Automatic bark recognition applied to tree species
classification is an important problematic that has gai-
ned interest in the computer vision community. It is
an interesting challenge to evaluate texture classifica-
tion algorithms on color images acquired in the wild.
Due to their low inter-class variability and high intra-
class variability, bark images are indeed considered as
difficult to classify for a machine as for a human.
In the context of tree species classification, barks
have interesting properties compared to other com-
monly used attributes (e.g. leaves, fruits, flowers,
etc.). Firstly, they are non-seasonal: Whatever the se-
ason, the texture of a bark will not change. This pro-
perty is particularly important to classify a tree during
winter, when no leaves or fruits are present. Secondly,
bark textures rarely change through short time periods
(i.e. a year basis), but they do change over long pe-
riods (i.e. tenth year basis). This property enables
the use of age priors for bark classification, but it ma-
kes the recognition even more challenging when these
priors are unavailable. Finally, barks are easier to iso-
late and to photograph compared to fruits and leaves
that may be unreachable on tall trees. In result, bark
images have been used to recognize tree species either
alone (Bertrand et al., 2017), or in combination with
other tree’s attributes (Bertrand et al., 2018). In or-
der to recognize trees, mobile applications like Folia
1
(Cerutti et al., 2013) have been developed. These ap-
plications assume that users do not necessarily have
an Internet connection, which is a very common si-
tuation in the wild. They should work on embedded
devices by seeking a trade-off between accuracy, time
complexity and space complexity to ensure state-of-
the-art recognition rates while avoiding unnecessary
energy consumption. Consequently, most of these ap-
plications are based on efficient handcrafted filters.
In this context, we propose two novel approaches
to drastically reduce the dimensionality of both tex-
ture and color feature vectors, while preserving accu-
racy. We focused our work on handcrafted methods
and decided not to use end-to-end Deep Convolutio-
nal Neural Networks (DCNNs) because existing da-
tasets are not well suited to both train and evaluate
deep learning algorithms: They contain relatively few
images, up to 1632 for 6 classes in (Porebski et al.,
2014). Moreover, DCNNs have both high time and
space complexity that make them unsuitable for em-
bedded usage.
To overcome the few numbers of segmented bark
classes in existing datasets; up to 12 in (
ˇ
Svab, 2014);
1
http://liris.univ-lyon2.fr/reves/folia/public/
240
Ratajczak, R., Bertrand, S., Crispim-Junior, C. and Tougne, L.
Efficient Bark Recognition in the Wild.
DOI: 10.5220/0007361902400248
In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 240-248
ISBN: 978-989-758-354-4
Copyright
c
 2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

Figure 1: Examples of bark images from Bark-101 dataset.
we also release a novel and even more challenging
dataset made of 101 segmented bark classes in high
resolution: Bark-101
2
(see Section 2.1). This dataset
has been conceived with a clear focus on the number
of classes, involving high intra-class and low inter-
class variabilities (see Figure 1).
The remainder of this study is organized as follow.
Section 2 presents previous work related to bark re-
cognition in the wild. Section 3 introduces and details
the proposed algorithms for efficient dimensionality
reduction of texture and color histograms. Section 4
presents the experiments we carried out and discusses
the results.
2 RELATED WORK
This section presents existing bark datasets and state-
of-the-art methods for bark recognition.
2.1 Datasets
We propose to describe four state-of-the-art datasets
that are publicly available and commonly used in the
literature. In complement, we release an even more
challenging dataset: Bark-101. The characteristics of
these datasets are detailed below and summarized on
Table 1.
BarkTex and NewBarkTex. BarkTex was the first
dataset specialized on bark recognition. It was intro-
duced by R. Lakmann (Lakmann, 1998). It is compo-
sed of 6 classes, each corresponding to a different tree
species. Each class contains 68 color images of trunks
for a total of 408 images. The trunks are spatially cen-
tered in the images, but some background can appear
depending on the width of the trunk. NewBarkTex is
derived from BarkTex. It was proposed in (Porebski
et al., 2014) and it is also composed of 6 classes. A
region of interest (ROI) of size 128x128 pixels was
cropped from the center of the images of BarkTex.
Then the ROI was separated in 4 sub-images of 64x64
pixels. Half of the sub-images were kept for training
and the second half for testing. Therefore, NewBark-
2
http://eidolon.univ-lyon2.fr/~remi1/Bark-101/
Tex is made of 272 images per class for a total of 1632
images.
Trunk12. It is a bark dataset created in 2014 by Matic
ˇ
Svab (
ˇ
Svab, 2014). It consists of 360 color images
of barks corresponding to 12 different species of tree
found in Slovenia. Each class consists of about 30
images. All images were taken with the same camera
in the same conditions (20 cm distance, multiple trees
per class, avoiding moss, same light conditions, taken
in upright position).
AFF. This bark dataset was presented in (Wendel
et al., 2011). It has 11 classes of bark from the most
common Austrian trees. It is composed of 1082 color
images of bark. In AFF, tree species are separated in
sub-classes according to the age of the tree. The tex-
ture of the trunk changes during the lifetime of a tree,
starting from a smooth to a more and more coarse tex-
ture. In this study, we have chosen to not separate the
classes according to the age of the trees.
Bark-101. We built the Bark-101 dataset among the
PlantCLEF
3
identification task, part of the Image-
CLEF challenge, designed to compare plant recogni-
tion algorithms. PlantCLEF contains photos of mul-
tiple yet not segmented plant organs (leaf, flower,
branch, steam, etc.) taken by people in various unsu-
pervised shooting conditions and gathered through
the mobile application Pl@ntNet
4
. More than 500
herb, tree and fern species centered on France are pre-
sent in PlantCLEF (Go
¨
eau et al., 2014). To construct
Bark-101, we kept the tree stems available from Plan-
tCLEF (i.e. the barks), and have manually segmen-
ted them to remove undesirable background informa-
tion. We decided to follow the authors of (Wendel
et al., 2011) by not constraining the size of the seg-
mented images in Bark-101. This choice was made
to simulate real world conditions assuming a perfect
stem segmentation. As a matter of fact, in a real world
setting not all trees would have the same diameter and
not all users would photograph them at a pre-defined
distance. The Bark-101 dataset is therefore compo-
sed of 101 classes of tree barks from various age and
size for a total of 2592 images. Images in Bark-101
contain noisy data like shadows, mosses or illumi-
3
https://www.imageclef.org/lifeclef/2017/plant
4
https://identify.plantnet-project.org/
Efficient Bark Recognition in the Wild
241

Table 1: Characteristics of bark datasets.
Dataset information BarkTex NewBarkTex Trunk12 AFF Bark-101
Classes 6 6 12 11 101
Total images 408 1632 393 1082 2592
Images per classes 68 272 30-45 16-213 2-138
Image size 256x384 64x64 1000x1334 1000x(478-1812) (69-800)x(112-804)
Illumination change 3 3 7 3 3
Scale change 3 3 7 3 3
Noise (shadows, lichen) 7 7 7 3 3
Train / Test splits 7 50/50 7 7 50/50
nation changes (see Figure 1). Due to the unsuper-
vised acquisition of the images and the variation of
bark textures over the lifespan of the tree, there is a
high intra-class variability in Bark-101. Furthermore,
a large number of classes naturally leads to a small
inter-class variability since the number of visually si-
milar species increases with the number of species.
Consequently, Bark-101 can be considered as a chal-
lenging dataset in the context of bark recognition. We
further demonstrate this statement in the experiments
presented in Section 4.
2.2 Existing Methods
Bark recognition is often considered as a texture clas-
sification problem. In (Wan et al., 2004) the aut-
hors compared different statistical analysis tools for
tree bark recognition, like co-occurrence matrices,
grayscale histogram analysis and run-length method
(RLM). The use of co-occurrence matrices on grays-
cale bark images is also present in (Huang et al.,
2006b). The authors combined them with fractal di-
mensions to characterize the self-similarity of bark
textures at different scales. Spectral methods, such as
Gabor filters, are also used. In (Huang et al., 2006a),
the authors demonstrated that only four wave orienta-
tions and 6 scales are sufficient to identify tree species
by their bark. To avoid losing the information provi-
ded by color, Wan et al. (Wan et al., 2004) applied
their grayscale method individually to each channel of
the RGB space. In (Baki
´
c et al., 2013), different color
spaces were used to characterize color information,
including RGB and HSV spaces. The authors of (Ber-
trand et al., 2017) proposed to combine texture and
color hues in a late fusion manner. First, they extrac-
ted orientation features using Gabor filters. Secondly,
they combined these features with a sparse represen-
tation of bark contours by encoding the intersections
of Canny edges with a regular grid. Lastly, they des-
cribed bark colors using the hue histogram from the
HSV color space. The resulting descriptor proved
to increase the classification rate of tree recognition
when combined with leaves (Bertrand et al., 2018).
Other commonly used descriptors for bark classi-
fication are Local Binary Pattern-like (LBP-like) des-
criptors that were inspired by the original LBP filter
proposed by (Ojala et al., 2001). LBP-like filters are
generic local texture descriptors parametrized over a
(P,R) neighborhood, with P the number of pixel neig-
hbors and R the radii. In the case of a multiscale fil-
ter, the number R
s
of radii R is strictly greater than 1.
LBP-like filters encode textural patterns with binary
codes, whose aggregation result in a texture histo-
gram of high dimension (e.g. R
s
× 2
P
for the original
LBP). Recently, the authors of (Boudra et al., 2018)
proposed a texture descriptor called Statistical Macro
Binary Pattern (SMBP) inspired by LBP. SMBP en-
codes the information between macrostructural scales
with a representation using statistical characteristics
for each scale. The descriptor increases the classifica-
tion rate on two barks datasets compared to the state
of the art. In (Alice Porebski, 2018), the authors de-
monstrated the accuracy of color LBP-like descriptors
with different color spaces, achieving above state-of-
the-art performance, but with a very high dimensio-
nal feature vector. Recent works on texture classifi-
cation in the wild are also of interest for bark recog-
nition. In particular, the Light Combination of Lo-
cal Binary Patterns (LCoLBP), proposed by (Rataj-
czak et al., 2019), and the Completed Local Binary
Pattern (CLBP), proposed by (Guo, Z. et al., 2010),
obtained equivalent results to popular DCNNs on his-
torical aerial images classification for a small compu-
tational cost. Since the CLBP filter is often used as a
baseline for bark recognition based on LBP-like filters
(Junmin Wang, 2017) the LCoLBP may be a suitable
candidate for bark recognition in the wild.
3 PROPOSED METHODS
Bark images acquired in the wild represent objects
with discriminative texture patterns and colors. To
represent these characteristics considering a trade-off
between accuracy and feature space complexity, we
propose two novel complementary methods to effi-
ciently reduce the number of texture and color fea-
tures. We considered these cues individually before
combining them in a late fusion manner (concatena-
tion). For the texture cues, we followed the com-
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
242







. . . .


 

 




  
  


 


 


 

 

  
1
0
mean
variance
entropy
max
min
median
kurtosis
1
0
. . .
1
0
. . .
. . .






  
80
16
32
48
64
80
16
32
48
64
-
80
16
32
48
64
. . . .
transpose & concatenate
. . . .
. . . .


 
Figure 2: Late Statistics (LS) from a Light Combination of Local Binary Patterns (LCoLBP) with a 3-radii neighborhood
(R
s
= 3), 8 neighbors P, and 7 statistics N
s
. Colored dots denote the statistics (outer colors) obtained from the N
h
= 5
sub-histograms included in the LCoLBP histogram for each radius (inner colors). This configuration yields a 2.3 reduction
factor.
mon use of LBP-like filters. Based on these filters,
we developed a generic yet efficient statistical repre-
sentation which preserves filters’ properties and does
not require re-sampling or mapping operations (see
Section 3.1). For the color cues, we were inspired
by (Bertrand et al., 2017) and built a task-guided low
dimensional histogram representation upon the HSV
color space using bark priors (see Section 3.2).
3.1 From Textures to Late Statistics
We define the Late Statistics as a combination of sta-
tistical features calculated from LBP-like texture his-
tograms. Considering a texture histogram H
t
, itself
made of the concatenation of N
h
known and orde-
red sub-histograms {h
1
,h
2
,...,h
N
h
}, we calculate N
s
Late Statistics independently for each h
i
with i ∈
{1,...,N
h
}. Late Statistics are then concatenated in
the same order as the h
i
sub-histograms. Assuming
a single H
t
per LBP-like radius, this process results
in a vector made of R
s
× N
s
× N
h
features, where R
s
is the number of radii (i.e. scales). It is represented
with N
s
= 7 statistics and a grayscale 3-radii (R
s
= 3)
LCoLBP filter on Figure 2. Note that each rotation of
the R-CRLBP (Ratajczak et al., 2018), a sub-filter of
the LCoLBP, is considered as an independent filter in
this study.
Late statistics are defined as late in opposition to
the early statistics proposed by (Boudra et al., 2018).
In (Boudra et al., 2018), the authors calculated sta-
tistics before calculating the texture histogram by re-
sampling the local textural patterns of a LBP-like fil-
ter. In consequence, the statistical approach proposed
by Boudra et al. would require new implementations
with error-prone re-sampling operations to be applied
to other LBP-like filters. The Late Statistics evade
this constraint by considering texture histograms that
have been already calculated. They do not need to
operate on the local textural patterns but rather on the
global histogram representations, so that they do not
require any modification of the descriptor implemen-
tations (i.e. no re-sampling).
Due to their nature, Late Statistics are expected to
preserve common descriptor properties, like rotation
and global illumination invariance. This point should
make the Late Statistics at least as robust to condition
changes as the descriptors themselves. Furthermore,
it should be observed that, similarly to the early statis-
tics used by (Boudra et al., 2018), the Late Statistics
naturally behave as a spatial normalization algorithm:
A histogram will be summarized with a fixed number
of statistics N
s
whatever its number of bins. This pro-
perty is particularly useful to combine different his-
tograms in a balanced feature vector that contains the
same quantity of information for each texture pattern.
Finally, one may observe that Late Statistics are
an efficient way to summarize textural information in
a very similar manner as Haralick features (Haralick
et al., 1973). However, while Haralick features repre-
sent statistics calculated directly from an image, Late
Statistics benefit from the efficient LBP-like represen-
tation of textures.
3.2 Efficient Colors
We discussed in Section 2 that several bark descrip-
tors work on grayscale images and that a few of them
also include color data. Among these descriptors, hue
information is commonly used (Baki
´
c et al., 2013;
Bertrand et al., 2017). The size of a hue histogram H
Efficient Bark Recognition in the Wild
243

0
b
add & shift
add & shift
0
Iteration 2
N-1 to N-2 bins
add & shift
0
Iteration 1
N to N-1 bins
Iteration 3
N-2 to N-3 bins
Output
N-3 bins
I


  

   


  

  

    

  


  

  

   

  
b
I
I
I
b
b

0
Figure 3: Schematic of the color histogram reduction applied with 3 iterations. The number of bins in the histogram is reduced
by 1 after each iteration. The reduction is performed using a local add and shift strategy on the bin b with the smallest intensity
I(b). A look-up table is used to map the modified bins.
is usually of 360 bins. However, this value (360) can-
not be stored on a single byte. For sampling reasons, it
is preferable to transpose the hue color range [0;359]
into the range [0;179] than into the range [0;255].
Therefore, the size of the hue histogram is 180 bins,
which represents a feature vector of a quite high di-
mensionality for task-specific applications with co-
lor priors, like bark recognition in the wild. As vi-
sible on Figure 1, barks have dominant variations of
brown, green and yellow colors. Based on this obser-
vation, we may expect that other colors like blue or
purple may not provide significant contribution to the
hue signature of a bark. However, completely remo-
ving these colors may result in less discriminant his-
tograms, making the bark classification process even
more difficult.
To reduce the size of the hue histogram, we pro-
pose to merge the least represented colors through a
non-destructive iterative process. For a given dataset
D made of k color images splitted into a train set T
r
and a test set T
e
, we first calculate and sum all the hue
histograms of the images in T
r
. We obtain a summed
hue histogram H
s
from the train set T
r
⊂ D. This sum-
med histogram H
s
is supposed to represent the color
prior on the whole dataset D. Secondly, on the sum-
med histogram H
s
, we iteratively add the population
of the bin b having the smallest intensity value (i.e.
the smallest population) to the population of its neig-
hbor of minimal population. After adding these bins,
we shift the histogram to the left in order to reduce its
dimension. The iteration process is stopped when the
desired size, fixed by the user, is reached (see Section
4). Attention should be brought to the circularity of
the hue channel in the HSV color space. The order
and position of the add and shift operations are stored
in a look-up table M. On the test set T
e
, the look-up ta-
ble M is then used to calculate the reduced histograms
of the input images L
T
e
by applying the add and shift
operations in the same order as in T
r
. Consequently,
the reduced histogram for an image l
T
e
∈ L
T
e
is gene-
rated in regard to the summed histogram, thus taking
into account the color prior of the data. The add and
shift process is illustrated on Figure 3.
4 EXPERIMENTS AND RESULTS
4.1 Experimental Setup
For our experiments, we considered two state-of-
the-art LBP-like filters made of complementary sub-
descriptors to assess the efficiency of the Late Statis-
tics: The Light Combination of Local Binary Patterns
(LCoLBP) and the Completed Local Binary Pattern
(CLBP). As explained in Section 2, these descriptors
are efficient texture filters which obtained DCNNs
like accuracy rates on texture in the wild datasets (Ra-
tajczak et al., 2019). Since these filters may result in
very high dimensional histograms, we considered a
constant number of neighbors P on a 3-radii (R
s
= 3)
neighborhood: (P = 8,R = {1,3, 5}). To assess the
effectiveness of the Late Statistics on both mapped
and unmapped LBP-like representations, we follo-
wed (Ratajczak et al., 2019) and applied the LCoLBP
without any mapping, resulting in a histogram of
R
s
×5 × 2
4
bins (N
h
= 5) with a (P = 8,R = {1,3,5})
neighborhood. On the other hand, we applied the ro-
tational and uniform (riu
2
) mapping introduced by
(Ojala et al., 2001) with the CLBP filter, resulting
in a histogram of 3 × (2 + 2 × 10) bins instead of
3 × (2 + 2 × 256) bins without mapping for the same
neighborhood.
We considered commonly used statistics in our ex-
periments: Mean, variance, entropy, minimum, maxi-
mum, median and kurtosis. Among them, we carried
out an ablation study through gridsearch to determine
the best combination of Late Statistics for each tex-
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
244

Table 2: Ablation study for the Late Statistics of the LCoLBP and CLBP filters on the BarkTex dataset.
Late Statistics Accuracy (%)
mean variance entropy minimum maximum median kurtosis LS-LCOLBP LS-CLBP
3 – – – – – – 81.9 71.8
3 3 – – – – – 82.8 59.6
3 3 3 – – – – 78.4 64.7
3 3 3 3 – – – 82.8 63.2
3 3 3 3 3 – – 83.1 69.4
3 3 3 3 3 3 – 86.3 72.1
3 3 3 3 3 3 3 89.5 62.8
3 3 – 3 3 3 3 88.2 60.1
3 – 3 3 3 3 3 89.5 62.5
3 – – 3 3 3 3 88.2 59.6
3 – – 3 3 3 – 88.2 75.3
ture filter as presented on Table 2 for BarkTex. Ab-
lation results obtained on other datasets were in con-
cordance with Table 2 as well as with the following
observations. On Table 2, we can observe that naively
adding more Late Statistics (first seven lines) may de-
crease the accuracy of the texture representations. We
also demonstrate that the Late Statistics should be ca-
refully and individually selected for each texture filter
in order to maximize the accuracy rate (up to a dif-
ference of
˜
15% on Table 2). This phenomenon high-
light a minor drawback of the Late Statistics: While
they are effective and easy to implement, they incre-
ase the number of parameters to tune. Therefore, the
number of statistics N
s
was set to 6 for the LCoLBP
and 4 to CLBP. We did not apply the Late Statistics on
the first sub-histogram of the CLBP filter because it is
made of only two bins. Late Statistics of the LCoLBP
filter (LS − LCoLBP) result in R
s
× 5 × N
s
features.
Late Statistics of the CLBP filter (LS −CLBP) result
in R
s
× (2 + 2 × N
s
) features.
Concerning the bark colors, we verified that bark
images are actually made of dominant colors on the
extensive train set of the Bark-101 dataset because
of the large variability of bark images and classes
available (see Section 2.1). The summed hue histo-
gram H
s
for this dataset is visible on Figure 4. Ba-
sed on this figure, we can observe that the majority of
barks present brownish, yellowish and greenish hues.
In order to define a suitable dimension for the redu-
ced color histogram, we observed the evolution of the
classification rate according to the strategy defined in
Figure 4: Summed histogram of the hue channels from the
train set of the Bark-101 dataset.
Section 4.2 on Bark-101 dataset (Figure 5). Despite
small variations in the species classification rate when
quantifying the histogram, we can see that the accu-
racy remains approximately constant over 30 bins. As
the other bark datasets present smaller variance and
are thus less representative of real world conditions,
we decided to set the size of the reduced hue histo-
gram to 30 bins regardless of the bark dataset. We
remind that a look-up table, to obtain the reduced his-
tograms, is calculated for each dataset independently:
Only the size of 30 bins has been fixed using Bark-
101.
4.2 Evaluation and Metrics
We separated our evaluation into two classification
strategies, c
1
and c
2
, depending on the dataset organi-
zations, visible on Table 1, and previous state-of-the-
art experiments. Strategy c
1
stands for the classical
train/test strategy. In c
1
, a part of the dataset is used
for training and the rest of the dataset is used for tes-
ting. We performed c
1
on NewBarkTex and Bark-101
following the train and test splits (50%/50%) provi-
ded by the authors. Strategy c
2
is the leave-one-out
strategy (LOO). In c
2
, we considered an ensemble S
of N samples and we performed N iterations. At each
iteration i ∈ {1,..., N}, we used a different sample s(i)
of S for evaluation (i.e. testing) and all the other sam-
ples of S −{s(i)} for training. If the evaluation sample
s(i) was successfully classified, the result of the cor-
responding iteration was set to 1, and to 0 otherwise.
The final accuracy was obtained by averaging the re-
sults of all iterations. In accordance with (Boudra
et al., 2018), we performed c
2
on BarkTex, Trunk12
and AFF. For both c
1
and c
2
we used the top-1 accu-
racy and a K-Nearest Neighbor classifier (KNN) with
K = 1 and the L1 distance. The 1-NN is the most com-
monly used classifier in the context of bark and tex-
ture recognition. The L1 distance has been chosen ar-
bitrarily. Additionally, for c
1
, and in accordance with
(Alice Porebski, 2018), we used a multi-class non-
linear Support Vector Machine (SVM) with radial ba-
Efficient Bark Recognition in the Wild
245

sis kernel. The parameters of the SVM classifier have
been optimized by grid search for each dataset and for
each feature vector.
4.3 Results and Discussion
Results are visible on Table 3 and Table 4. They
have been obtained considering the evaluation stra-
tegies described in Section 4.2. Highest accuracy ra-
tes from other studies have been reported and marked
with a right-top star symbol. Specifically, for AFF,
Trunk12 and BarkTex, we reported the results obtai-
ned with MSLBP* and SLBP* from (Boudra et al.,
2018). The results for NewBarkTex were reported for
Wang17* (Junmin Wang, 2017), Sandid16* (Sandid
and Douik, 2016), and Porebski18* (Alice Porebski,
2018). We also considered the results proposed by the
methods of (Bertrand et al., 2017) that we renamed
GWs and GWs/H180. All other results correspond to
our own implementations using OpenCV 3.4 in C++
for the texture and color descriptors. Scikit-learn in
Python has been used for the Late Statistics and the
classifiers. The texture cues were calculated on grays-
cale images. The color cues were calculated on bark
images in HSV color space. The following sections
discuss the results obtained.
4.3.1 Color Cues
On Tables 3 and 4, we can observe that all the tex-
ture filters obtained higher accuracy rates when com-
bined with color histograms of both 30 bins (H30)
and 180 bins (H180). As a reminder, H30 is the re-
duced hue histogram and H180 is the complete hue
histogram. When used alone, H180 is, in average,
only 3.3% more accurate than H30 on AFF, Trunk12
and BarkTex. However, we found that when combi-
ned with grayscale textures, the contribution of both
color representations is equivalent. These results de-
monstrate the efficiency of the color histogram re-
duction algorithm presented in Section 3.2. Moreo-
Figure 5: Accuracy of the reduced hue histogram on the
Bark-101 dataset depending on the final number of bins.
Table 3: Results of Late Statistics and reduced color histo-
grams with KNN and leave-one-out strategy. Blue: Highest
results of the literature. Green: highest results overall. Red:
Highest late statistics results.
Top-1 Accuracy / Dataset (%)
Descriptor size
AFF Trunk12 BarkTex
MSLBP* 2 816 63.3 63.3 86.8
SMBP* 10 240 71.7 71.0 84.3
H30 30 50.5 64.4 55.4
H180 180 55.6 69.0 61.3
LCoLBP 240 75.3 77.1 92.1
LCoLBP / H30 270 80.7 84.2 92.4
LCoLBP / H180 420 80.7 84.2 91.7
CLBP 66 68.1 70.0 78.7
CLBP / H30 96 72.9 77.4 83.8
CLBP / H180 246 73.5 78.1 84.3
GWs 121 48.2 39.9 56.1
GWs / H30 151 64.7 74.3 66.2
GWs / H180 301 66.5 76.1 69.6
LS-LCoLBP 90 69.4 74.6 89.5
LS-LCoLBP / H30 120 76.9 80.7 90.2
LS-LCoLBP / H180 270 76.9 80.7 91.2
LS-CLBP 30 59.1 70.0 75.3
LS-CLBP / H30 60 65.4 77.4 78.2
LS-CLBP / H180 210 67.9 78.1 79.4
Table 4: Results achieved on NewBarkTex and Bark-101.
Top-1 Accuracy / Dataset(%)
NewBarkTex Bark-101
Descriptor size KNN SVM KNN SVM
Porebski18* 10 752 – 92.6 – –
Wang17* 267 84.3 – – –
Sandid16* 3 072 – 82.1 – –
H30 30 48.0 50.6 19.1 20.4
H180 180 48.5 53.6 22.2 20.9
LCoLBP 240 78.8 89.3 34.2 41.9
LS-LCoLBP 90 66.5 79.4 28.3 30.1
LS-LCoLBP / H30 120 71.9 82.0 27.6 32.1
LS-LCoLBP / H180 270 72.3 82.2 27.8 31.0
GWs / H30 151 60.4 74.1 28.2 31.7
GWs / H180 301 54.1 63.6 31.8 32.2
ver, these results confirm that color cues seem to be
non-negligible features for bark recognition, in oppo-
sition to the assumption made by (Boudra et al., 2018)
but in accordance with (Junmin Wang, 2017).
4.3.2 Late Statistics
The Late Statistics decreased the size of the evaluated
texture features with a multiplicative factor between
2.7 for the LCoLBP and 2.2 for the CLBP with an
averaged reduction in accuracy of only 5.5% overall.
The Late Statistics seem particularly efficient with a
leave-one-out strategy (Table 3) with an averaged dif-
ference of 3.7% between the LCoLBP and the LS-
LCoLBP, and an averaged difference of 4.1% between
the CLBP and the LS-CLBP. On the other hand, they
are slightly less effective on the Train/Test strategies.
These results may be explained by a lack of training
data resulting in an under accurate statistical descrip-
tion of the per-class texture histograms.
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
246

4.3.3 Overall Performances
We can observe that the Late Statistics combined with
the reduced hue histograms H30 outperform prior
works on the AFF, Trunk12 and BarkTex datasets.
LS-LCoLBP/H30 is in averaged 6.3% more accurate
than the methods compared in (Boudra et al., 2018).
Moreover, it is about 100 times smaller than SMBP*.
On NewBarkTex, Late Statistics combined with hue
histograms and a SVM classifier achieve similar re-
sults to Sandid16* with an accuracy of 82.0%. We
observe that our method (LS − LCoLBP/H30) obtai-
ned slightly lower results than the most accurate algo-
rithms from the literature on this dataset, but it does
have a significantly smaller feature vector which is
about 30 times smaller than Sandid16* and 100 times
smaller than Porebski18*. On Bark-101, we can ob-
serve the lowest accuracy rates for all compared met-
hods over all the datasets. These results are explained
by the higher number of classes in Bark-101 compa-
red to existing datasets. It also demonstrates the chal-
lenge proposed by Bark-101. However, most methods
including GWs/H30 achieved a top-1 recognition rate
about 30%, which is far above the random guess of
0.9%.
5 CONCLUSION
In this study, we compared recent state-of-the-art des-
criptors in the context of tree bark recognition in the
wild. We proposed two novel algorithms to reduce
the dimensionality of texture and color features vec-
tors. We showed that the proposed algorithms out-
perform state-of-the-art methods on four bark datasets
with a considerable gain in space complexity. We be-
lieve that these methods can be generalized on other
histogram-like feature vectors. Furthermore, we rele-
ased a new dataset made of 101 bark classes of seg-
mented images with high intra-class variability. We
demonstrated that the proposed dataset is particularly
challenging for existing methods, enforcing the need
for future prospects on bark recognition. Future work
will investigate the proposed methods as a lightweight
representation with multiple color spaces. We will
also evaluate the proposed algorithms on mobile plat-
forms, such as smartphones, to assess their perfor-
mances on real-world settings.
ACKNOWLEDGEMENT
This work is part of ReVeRIES project (Reconnais-
sance de V
´
eg
´
etaux R
´
ecr
´
eative, Int
´
eractive et Educa-
tive sur Smartphone) supported by the French Nati-
onal Agency for Research with the reference ANR-
15-CE38-004-01, and part of the French Environment
and Energy Management Agency, Grant TEZ17-42.
REFERENCES
Alice Porebski, Vinh Truong Hoang, N. V. D. H. (2018).
Multi-color space local binary pattern-based feature
selection for texture classification. Journal of Electro-
nic Imaging, 27:27 – 27 – 15.
Baki
´
c, V., Mouine, S., Ouertani-Litayem, S., Verroust-
Blondet, A., Yahiaoui, I., Go
¨
eau, H., and Joly, A.
(2013). Inria’s participation at imageclef 2013 plant
identification task. In CLEF Working Notes.
Bertrand, S., Ameur, R. B., Cerutti, G., Coquin, D., Valet,
L., and Tougne, L. (2018). Bark and leaf fusion sy-
stems to improve automatic tree species recognition.
Ecological Informatics.
Bertrand, S., Cerutti, G., and Tougne, L. (2017). Bark re-
cognition to improve leaf-based classification in di-
dactic tree species identification. In VISAPP.
Boudra, S., Yahiaoui, I., and Behloul, A. (2018). Plant iden-
tification from bark: A texture description based on
statistical macro binary pattern. In ICPR.
Cerutti, G., Tougne, L., Mille, J., Vacavant, A., and Coquin,
D. (2013). Understanding leaves in natural images - a
model-based approach for tree species identification.
Computer Vision and Image Understanding.
Go
¨
eau, H., Joly, A., Bonnet, P., Selmi, S., Molino, J.-
F., Barth
´
el
´
emy, D., and Boujemaa, N. (2014). Li-
feclef plant identification task 2014. In CLEF2014
Working Notes. Working Notes for CLEF 2014 Con-
ference, Sheffield, UK, September 15-18, 2014, pages
598–615. CEUR-WS.
Guo, Z., Zhang, L., and Zhang, D. (2010). A Completed
Modeling of Local Binary Pattern Operator for Tex-
ture Classification. IEEE TIP, 19(6).
Haralick, R. M., Shanmugam, K., and Dinstein, I. (1973).
Textural Features for Image Classification. IEEE
Transactions on Systems, Man, and Cybernetics.
Huang, Z.-K., Huang, D.-S., Du, J.-X., Quan, Z.-H., and
Guo, S.-B. (2006a). Bark classification based on gabor
filter features using rbpnn neural network. In Interna-
tional conference on neural information processing,
pages 80–87. Springer.
Huang, Z.-K., Zheng, C.-H., Du, J.-X., and Wan, Y.-y.
(2006b). Bark classification based on textural featu-
res using artificial neural networks. In International
Symposium on Neural Networks. Springer.
Junmin Wang, Yangyu Fan, N. L. (2017). Combining fine
texture and coarse color features for color texture clas-
sification. Journal of Electronic Imaging.
Lakmann, R. (1998). Barktex benchmark database of color
textured images. Koblenz-Landau University.
Ojala, T., Pietik
¨
ainen, M., and M
¨
aenp
¨
a
¨
a, T. (2001). A Gene-
ralized Local Binary Pattern Operator for Multiresolu-
tion Gray Scale and Rotation Invariant Texture Classi-
Efficient Bark Recognition in the Wild
247

fication. In Advances in Pattern Recognition, ICAPR
2001, pages 399–408. Springer Berlin Heidelberg.
Porebski, A., Vandenbroucke, N., Macaire, L., and Hamad,
D. (2014). A new benchmark image test suite for eva-
luating colour texture classification schemes. Multi-
media Tools and Applications, 70(1):543–556.
Ratajczak, R., Crispim-Junior, C., Faure,
´
E., Fervers, B.,
and Tougne, L. (2019). Automatic land cover recon-
struction from historical aerial images: An evaluation
of features extraction and classification algorithms.
IEEE TIP (accepted with minor revision).
Ratajczak, R., Crispim-Junior, C. F., Faure,
´
E., Fervers,
B., and Tougne, L. (2018). Reconstruction automati-
que de l’occupation du sol
`
a partir d’images a
´
eriennes
historiques monochromes : une
´
etude comparative.
In Conf
´
erence Franc¸aise de Photogramm
´
etrie et de
T
´
el
´
ed
´
etection (CFPT), Marne-la-Vall
´
ee, France.
Sandid, F. and Douik, A. (2016). Robust color texture des-
criptor for material recognition. PRL.
ˇ
Svab, M. (2014). Computer-vision-based tree trunk recog-
nition. PhD thesis, Fakulteta za raunalnitvo in infor-
matiko, Univerza v Ljubljani.
Wan, Y.-Y., Du, J.-X., Huang, D.-S., Chi, Z., Cheung, Y.-
M., Wang, X.-F., and Zhang, G.-J. (2004). Bark tex-
ture feature extraction based on statistical texture ana-
lysis. In International Symposium on Intelligent Mul-
timedia, Video and Speech Processing. IEEE.
Wendel, A., Sternig, S., and Godec, M. (2011). Automated
identification of tree species from images of the bark,
leaves and needles. In 16th Computer Vision Winter
Workshop.
VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications
248