Feature Extraction, Feature Selection and Machine
Learning for Image Classification: A Case Study
Mădălina Cosmina Popescu1 , Lucian Mircea Sasu2 ,3
1 Faculty of Mathematics and Computer Science, Transilvania University of Braşov, Romania, cosmina.popescu@xu.unitbv.ro
2 Mathematics and Computer Science Department, Transilvania University of Braşov, Romania, lmsasu@unitbv.ro
3 Siemens Corporate Technology RTC, Braşov, Romania
Abstract—This paper presents feature extraction, feature se-
lection and machine learning–based classification techniques for
pollen recognition from images. The number of images is small
compared both to the number of derived quantitative features
and to the number of classes. The main subject is investigation of
the effectiveness of 11 feature extraction/feature selection algo-
rithms and of 12 machine learning–based classifiers. It is found
that some of the specified feature extraction/selection algorithms
and some of the classifiers exhibited consistent behavior for this
dataset.
I. I NTRODUCTION
In palynology, a specific task is recognizing plants from
the pollen grains. Although domain expertise as expressed by
highly qualified palynologist is for now the standard approach,
we witness an increasing interest in applying automatic learn-
ing approaches for plant recognition.
There are at least two approaches which can be used for
automatic image classification. The former is the classical
machine learning workflow for image classification: a) use
some chain of image-processing techniques to extract quan-
titative features from each image; b) apply automatic feature
extraction/selection techniques; c) apply some classifiers on
the data subsets. The mix between the considered numerical
features and classifiers dictates the success of the recognition
task.
During feature extraction, a human expert might decide to
follow or to avoid some of the processing techniques. For
classification models, one can find scarce prior information
on which of them might be successful, based solely on the
task’s particularities. We recall Tom Mitchell’s statements that
any useful machine learning algorithm must have its own in-
ductive bias and that “progress towards understanding learning
mechanisms depend upon understanding the sources of, and
justification for, various biases”, beyond the particularities of
the training sets [1]. Despite many theoretical results and
experimental analysis of machine learning models, machine
learning largely remains an experimental science.
The latter approach used for automatic image classification
is a more recent one, and allows the machine to extract features
on its own, without human intervention. We refer here to the
so-called “deep learning” branch of machine learning [2]–[5].
Computer vision is the most popular choice as a testbed for
deep learning architectures: the first successful story is the
usage of convolutional neural networks for optical character
978-1-4799-5183-3/14/$31.00 ' 2014 IEEE 968
recognition with minimal preprocessing [6]; Krizhevsky et
al. [7] use a neural network to classify 1.2 million high
resolution images, reaching 37.5% top-1 error rates, much
better than previous results; in [8], Kavukcuoglu et al. propose
an unsupervised method for learning multi–stage hierarchies of
sparse convolutional features, improving performance on some
visual recognition and detection tasks. A machine learning
classification system which outruns human performance in
traffic sign recognition is described by Ciresan et al. in [9]
and for handwritten digit recognition in [10].
The present paper follows the classical machine–learning
workflow. Related work was performed by Langford, Taylor
and Flenley [11] and they obtained 94.3% correct classification
on a 6 pollen class database with 192 samples. The images
were obtained through scanning electron microscope, an ex-
pensive and specialized solution. For a 3–class dataset – also
provided by scanning electron microscope – an artificial neural
network was able to perform 100% accuracy [12]. Getting
the images through electronic scanning is a rather exclusivist
method, and further researchers considered regular images, i.e.
images captured by normal light-sensitive cameras from an
optical microscope.
In [13] 90% classification accuracy was obtained for a 3-
class problem. Zhang et al. [14] reported 91% classification
success using wavelet decomposition and reconstruction and
combination with coocurrence matrices, followed by learning
vector quantization and backpropagation networks. Dell’Anna
et al. [15] used a k–nearest neighbors classifier starting
from data provided by Fourier transform infrared microspec-
troscopy; the discrimination between the 11 taxonomic groups
was accomplished with 84% accuracy.
For some of these reports we remark the small number
of classes and explicit mention of small data sets, finalized
with large percent of correct classification. It seems to be
an inverse correlation between the number of classes and the
classification performance: for example, Boucher and Thonnat
[16] obtain 77% and 97% correct classification for 30 and 4
classes, respectively.
In our study we started from a public pollen image dataset1 .
The challenge of this problem is given by the small number
of available images – between 4 and 7 for each of the 38
1 The set of images can be downloaded from
http://www.geo.arizona.edu/palynology/sem/nucastl.html.
species considered, with an average of about 4.76 images per
species. The scarcity of the data makes it prone to overfitting
during classification model development. Additionally, the
number of the numerical attributes (48) derived through image
processing techniques is large compared to the number of
available records. Considerable effort was put into feature
extraction and selection, and multiple classification models
were considered. An assessment process based on stratified
k–fold cross validation, data shuffling and t-tests was finally
considered in order to rank the classifiers. It turns out that
some feature extraction/selection algorithms and some of the
classifiers exhibit consistent behavior, and this finding is
a key contribution of the paper. Beside this, the reported
classification scores could serve as a benchmark for further
attempts on this dataset.
The outline of the paper is as follows: Section II contains a
description of the dataset we used in our study and extraction
of quantitative features from images. Section III describes the
feature extraction and feature selection preprocessing methods,
the classification models and the performance assessment
procedure. Section IV contains the experimental results and
comments on the feature extraction/selection effectiveness and
the classification performance for the machine learning models
considered. Section V concludes the paper.
II. DATASET DESCRIPTION AND QUANTITATIVE FEATURES
EXTRACTION
A. Dataset Description
In our study, we used a publicly available dataset from the
Pollen Laboratory of The University of Newcastle, Australia.
From the mentioned dataset, we selected only those species
for which at least 4 images were available at the time when
we started this research. The reason for this selection is that,
in classification, we cannot obtain relevant results using 1 to
3 instances per class. We used 38 species of plants collected
in a total of 181 images. There are 19, 10, 8, and 1 species
depicted by 4, 5, 6, and 7 images, respectively. This gives an
average of 4.76 images per species.
In order to derive image classifiers, we extracted 48 nu-
merical attributes from each pollen grain image. This was
accomplished in two steps: image enhancement and numerical
attribute extraction.
B. Image Enhancement for Foreground Extraction
The computation of attributes, as it is presented in the
following section, requires to first compute the pollen grain
mask, i.e. an image where the foreground is represented by
white pixels and the background by black pixels. The original
image is then masked, in order to separate foreground from
background in pollen grain image. The preprocessing steps are
described below.
First, the image is stretched, to reduce autocorrelation. The
image is split into R, G, B channels which are stretched
separately, then the inverse of splitting operation is applied
to put the channels back together. The resulted image is
transformed from RGB into HSV color space, in order to
969
use the saturation channel; this step helps us to separate
foreground from background, since in the pollen grain image
the saturation of foreground is higher than the saturation of
background.
Further, the contrast is increased by equalizing the his-
togram for saturation channel and applying gamma-correction
with γ = 3 [17]. We chose this value for gamma firstly because
gamma-correction with γ > 1 exponentially increases intensity
for high-intensity pixels, and decreases it for low-intensity
ones. The exact value is chosen experimentally, to fit our set
of images. Next, an adaptive threshold is applied, to obtain a
binary image. The threshold is computed for each image using
data from image histogram [17]. Our set of images required
to apply a Gaussian filter before thresholding, to make this
operation less sensitive to noise.
At this point, the binary image that we obtained contains
black holes within. To eliminate them, we applied the fol-
lowing set of morphological operations: dilation, hole filling,
erosion, and extraction of central connected component [17].
C. Numerical Attribute Extraction
This step starts from the processed images obtained in
section II-B and produces numerical descriptors, which can
be further used as input values in machine learning based
classifiers.
We use the attributes presented in [18]: area of the pollen
grain, width and height of the bounding box, coordinates x
and y of the grain centroid, major axis length and minor
axis length of the ellipse which has the same second moment
as the pollen shape, area of the convex hull enclosing the
shape, diameter of the circle with the same area as the object,
solidity, perimeter of the mask image, extent, eccentricity,
coordinates x and y of the weighed centroid, shape, thickness,
box (x and y coordinates for both upper-left and bottom-right
corners of an inner rectangle of the bounding box), height,
width, correlation, homogeneity, contrast, entropy (computed
for 6 images: the blue channel of the RGB color space, the
saturation and value channels from HSV color space, each of
them extracted from the bounding box and the box), Fourier
Descriptors (the first 5 Fourier coefficients, as described below
in this section), relative areas (5 values, computed from the
binary image of the pollen grain, as presented in [18]), and
relative objects (5 values based on the images obtained in
relative areas computation, as explained at the end of this
section).
Some of the mentioned attributes are clearly explained in
the cited work and we will not go into details for them. To
compute the values for the remaining attributes (the ones that
are briefly described in [18]) additional research was required.
Below we provide some explanation of the algorithms behind
attribute computation.
For the computation of major axis length and minor axis
length of the ellipse which has the same second moment as
the pollen shape we use the covariance matrix of white pixels
coordinates. Calculating the eigenvectors of this matrix, we
obtain the pixel distribution orientation. The minor axis length
 We had to represent the contour applying discrete Fourier
transform to complex numbers given by pixel coordinates.
From the resulted coefficients we used only the first 5 values,
as suggested in [18]. In our vector of features, we use the
magnitude of the 5 selected descriptors.
Relative objects are represented by numbers of connected
components in inverted and masked version of the binarized
(a) Original image (b) Stretched (c) Saturation chan- images computed for relative areas. To obtain the connected
nel components from the binarized image we used the Two-Pass
Connected Components Labeling algorithm [19].
As a result of these steps, we obtain 48 numerical descrip-
tors for each image, and we call this dataset “pollen”.
III. E XPERIMENTAL S ETUP
A. Feature extraction and selection
One legitimate question is whether all these 48 features
(d) Histogram (e) Gamma correc- (f) Gaussian filter are suitable for the classification task, or whether a lower
equalization tion cardinality set of features would improve the classification.
The existing literature describes an impressing number of
approaches devoted to feature selection and feature extraction
– see for example [20]–[25]. The implementations we used in
this paper are provided by Weka [26].
One popular feature extraction technique is based on prin-
cipal component analysis; it is the only feature extraction
method we consider here. For our dataset we chose enough
(g) Adaptive (h) Dilation (i) Hole filling eigenvectors to account for 95% of the variance in the original
threshold data, and thus 17 derived features were produced; we hence
get the “pca” dataset.
Another candidate preprocessing step is lowering the num-
ber of linearly correlated features. It is known that highly
correlated features rise unnecessary computational burden and
worsens the classification accuracy for most linear and for
few nonlinear techniques [25]. The resulted dataset is denoted
as “uncorr” and contains 34 features whose pairwise absolute
(j) Erosion (k) Central (l) Foreground sep- correlation value is at most 95%.
connected aration The next nine feature selection methods share the following
component
strategy: the value of a subset of attributes is evaluated by
Fig. 1. Image enhancement steps for foreground extraction, applied to Calytrix counting the predictive ability of each feature along with the
tetragona pollen grain (in common language, fringe-myrtle).
degree of redundancy between them.
The considered heuristics are: best-first, which either starts
with an empty set of candidate features which is succesively
and major axis length are the eigenvalues corresponding to grown – the so-called forward search – or with the set of all
the above mentioned eigenvectors. Eccentricity quantifies how available features from which it performs eliminations – the
much the shape deviates from being circular. backward search. The resulted datasets are called “bff” and
Height is the length of the largest line enclosed in the “bfb”, respectively. The third heuristic is based on evolutionary
pollen. We computed this as the maximum length of a line algorithms; the resulted dataset is named “ea”. The fourth
that passes through 2 points on the contour and the centroid. strategy performs the search based on genetic algorithm [27]
Width is the length of the largest line enclosed in the pollen and produces the dataset “ga”. The fifth strategy performs a
and perpendicular to Height. greedy search in the space of attributes, augmenting the initial
By using Fourier descriptors we translated points of the empty set of features until the performance starts to decrease;
shape contour from spatial domain to frequency domain and the corresponding dataset is named “gs”. The linear forward
extracted only the terms corresponding to low frequencies. selection is an extension to the best-first approach [28]; we
Unlike the terms that correspond to high frequencies, they give denote the corresponding dataset as “lfs”. The seventh feature
information about the shape, ignoring details such as small selection strategy employs particle swarm optimization [29]
oscillations on the contour. This aspect is of real importance as an exploration method and produces the dataset “pso”. The
for generalization. eighth selection strategy – reranking search – is an information

970
theoretic based meta-search algorithm proposed in [30]; the
resulted dataset is denoted as “rrs”. The last heuristic is tabu
search [31] and produces the dataset “ts”.
After reducing the features count according to the proce-
dures described above, the number of input features ranged
between 15 and 34. We also consider the set of all 48 features,
to check whether the feature selection/extraction procedures
impact the classification accuracy. In a nutshell, we have 12
datasets which are in turns used for the classification step2 .
B. Classification models
The machine learning literature contains a large number
of classification approaches. Unfortunately, it is also scarce
when looking for advice on how to properly choose among
them. Although there are some attempts to formulate some
strategies from experimental results (e.g. [32]) – the machine
learning related research is nowadays tributary to trial–and–
error approach.
The following 12 classification methods were investigated
in turn: naive Bayes (NB), multinomial logistic regression
(MLR), multilayer perceptron (MLP), radial basis function
network (RBF), k-nearest neighbors (KNN), the rule–based
classifier PART [33], functional trees (FT) [34], [35], best-first
decision tree (BFT) [36], Hoeffding tree (HT) [37], logistic
model trees (LMT) [35], C4.5 [38], and Random forest (RF)
[39].
Some of the hyperparameters of the models were tuned,
based on trial and error. For example, in case of KNN, we used
leave-one-out strategy to decide which value k ∈ {1, . . . , 20}
gives the best result; for the multilayer perceptron, through
cross validation we searched for the number of hidden neurons
and the learning rate, while keeping a constant momentum
coefficient.
C. Performance assessment
The performance assessments were accomplished using the
Weka Experiment Environment [26]. Each of the 12 datasets
was randomly shuffled 10 times. For each shuffle and each of
the 12 classification models its percent of correct classification
(PCC) was averaged using 4-fold stratified cross validation.
The number of folds is dictated by the classes distribution:
in the considered dataset, there are 19 species for which
only 4 pollen grain images were available. The differences
between the results were quantified using statistical methods
on the 12 datasets · 10 shuffles · 4 folds · 12 classifiers = 5760
measurements.
IV. E XPERIMENTAL RESULTS
A. Feature extraction/selection influence
We start by discussing the influence of the feature extrac-
tion/selection strategies on classification results. One remarks
a consistent behavior across the various classifiers for some of
these strategies. In 8 cases out of 12, the greedy search leads
to the worst classification results. The second and the third
2 These arff datasets are available for download
http://cs.unitbv.ro/∼lmsasu/data/pollen.arff.
lowest PCCs are produced by the best–first with backward
search strategy and principal component analysis (8 cases
each), respectively. A noteworthy remark is that the three
mentioned strategies produce all the lowest three PCC results
for all classifiers. For the numerical features we got from
image processing, it clearly signifies that greedy search, best–
first with backward search and principal component analysis
must be avoided as feature extraction/selection strategies.
For the best influence on the classification step, the results
are not so well grouped: the largest PCC values were obtained
by evolutionary algorithms (5 cases out of 12), reranking
search, linear forward selection, and tabu search (each with 3
cases out of 12). Note that there are some ties: for multilayer
perceptron, best-first with forward search ties with linear
forward selection and tabu search; for functional trees, the best
score is shared by best-first with forward search, evolutionary
algorithms, linear forward selection and tabu search.
The second largest PCCs are obtained for evolutionary
and genetic algorithm (4 cases each), best-first with forward
search, linear forward selection, and tabu search (2 cases
each). There are ties between best-first with forward search,
linear forward selection and tabu search for Hoeffding tree and
logistic model trees.
The third largest PCC is obtained by reranking search (5
cases), best-first with forward search, genetic algorithm, linear
forward selection, and tabu search (3 cases each). There are
ties between best-first with forward search, linear forward
selection and tabu search for multilayer perceptron, k-nearest
neighbors and random forests.
In short, when we refer to the highest 3 PCCs (ties
included), evolutionary algorithms and reranking search are
encountered 9 times, the best-first with forward search and
linear forward selection appears 8 times, and genetic algorithm
appears 7 times.
As a conclusion on the effect of feature extraction/selection,
the worst behavior is given exclusively by greedy search, best–
first with backward search and principal component analysis;
the highest PCCs are found for evolutionary and genetic
algorithms, reranking search, best-first with forward search,
and linear forward selection.
B. Classifiers’ performances
For classifiers comparisons we took naive Bayes as a
baseline model, as its PCC is close to the median PCC
over all datasets. Starting from the 5760 measurements, we
used corrected resampled t-test [40] as implemented in the
Weka Experimental Environment. Table I marks the relation
between naive Bayes and the other classifiers as follows: a
“(v)” annotation shows that a result is statistically better than
the baseline model; “(*)” marks the results that are statistically
worse than the baseline model; all unannotated results are
statistically the same as the one obtained by NB. The two
summary rows contain x/y/z triples showing how many times
from each classifier was (x) statistically better than, (y) the same
as, and respectively (z) worse than the baseline model.
971
 TABLE I
E XPERIMENTAL RESULTS OF PCC FOR THE 12 DATASETS AND 12 CLASSIFIERS USING STRATIFIED 4- FOLD CROSS VALIDATION AND 10 RANDOM
SHUFFLES .

Dataset NB MLR MLP RBF KNN PART

pollen 63.99 66.73 77.35 (v) 74.98 (v) 76.09 (v) 44.91 (*)
pca 52.42 54.44 65.83 (v) 56.78 58.00 39.28 (*)
uncorr 64.71 68.34 77.46 (v) 74.43 (v) 73.83 (v) 43.76 (*)
bff 67.25 77.91 (v) 80.94 (v) 77.03 (v) 82.66 (v) 42.28 (*)
bfb 48.12 53.43 56.25 58.23 (v) 62.76 (v) 39.46
ea 68.30 76.29 (v) 76.84 (v) 78.74 (v) 83.21 (v) 46.07 (*)
ga 68.23 78.83 (v) 80.78 (v) 77.46 (v) 81.16 (v) 46.53 (*)
gs 46.46 50.34 54.53 54.96 (v) 60.62 (v) 40.12
lfs 67.25 77.91 (v) 80.94 (v) 77.03 (v) 82.66 (v) 42.28 (*)
pso 64.93 73.64 (v) 74.37 (v) 79.12 (v) 79.73 (v) 44.81 (*)
rrs 67.26 79.77 (v) 80.27 (v) 76.86 (v) 83.43 (v) 45.53 (*)
ts 67.25 77.91 (v) 80.94 (v) 77.03 (v) 82.66 (v) 42.28 (*)
Summary 7/5/0 10/2/0 11/1/0 11/1/0 0/2/10
Dataset FT BFT HT LMT C4.5 RF
pollen 75.26 (v) 45.81 (*) 59.85 74.64 49.84 (*) 62.14
pca 63.41 (v) 40.51 52.53 61.37 40.50 (*) 50.77
uncorr 76.51 (v) 47.68 (*) 59.45 76.45 (v) 50.23 (*) 63.58
bff 82.05 (v) 46.56 (*) 66.42 81.50 (v) 47.15 (*) 64.08
bfb 56.14 39.50 46.19 53.59 44.26 48.35
ea 82.05 (v) 49.65 (*) 67.08 81.88 (v) 48.90 (*) 64.36
ga 81.43 (v) 49.61 (*) 64.87 80.77 (v) 50.61 (*) 63.75
gs 54.87 39.50 45.58 53.98 43.86 46.52
lfs 82.05 (v) 46.56 (*) 66.42 81.50 (v) 47.15 (*) 64.08
pso 78.56 (v) 47.71 (*) 62.00 78.94 (v) 47.75 (*) 62.16
rrs 81.83 (v) 44.25 (*) 66.41 80.89 (v) 48.48 (*) 65.70
ts 82.05 (v) 46.56 (*) 66.42 81.50 (v) 47.15 (*) 64.08
Summary 10/2/0 0/3/9 0/12/0 8/4/0 0/2/10 0/12/0

The performance, as quantified by the x/y/z triples natu-
rally clustered in four groups. The top four performance was
given by RBF, KNN (each having (x/y/z) = (11/1/0)), MLP
and FT (both with (x/y/z) = (10/2/0)). The RBF and KNN
– the two with highest x values – are both local models.
The only dataset for which RBF and KNN did not provide
statistically better PCCs than NB is the one obtained through
principal component analysis. MLP and FT underperformed
on the “pca” and “gs” datasets. It is interesting to note that
these four classifiers were the only ones for which the original
dataset “pollen” provided statistically better PCCs than NB.
The next performance group contains LMT and MLR,
providing statistically better (the same) behavior as NB in 8
(4) and 7 (5) cases, respectively. Hence, one can consider they
are still better than the baseline classifier, although at a lower
margin than the previous cluster. There is an almost perfect
overlap of the datasets for which the LMT and the MLR
classifiers outperform NB. For the original “pollen” dataset the
LMT, MLR and the NB models lead to statistically identical
scores.
The HT and RF were statistically equivalent to naive Bayes
for this problem for all datasets. They give the third cluster of
performance.
Finally, the weakest performance was obtained by PART,
BFT and C4.5; in at least 9 of the data subsets their per-
formance was statistically inferior to NB, and they never
performed better than it.
A remark on the efficiency of the original “pollen” dataset:
as seen in Table I, in 4 cases it produces statistically higher
PCCs than the baseline model; this happens for the 4 classi-
fiers which compose the highest performance cluster, so the
results might be due to the robustness of these models. The
performance is counterbalanced by 3 cases when they produce
statistically lower PCCs than NB (PART, C4.5 and BFT). This
later result might be explained by the overall poor performance
of the involved classifiers.
The PCC scores might appear as modest at first sight: only
few of the results are beyond 80% PCC. The best results
mentioned in the introduction are obtained for a quite small
number of classes, while in our case, the classifiers deal with
38 species (classes). The scarcity of the dataset – an average
of 4.76 images/species – is another potential cause of these
quite modest PCC values.
V. C ONCLUSIONS
This paper started from a commonly encountered task in
palynology: classification of plants using pollen grain images.
The aim is to automate this task, from image processing to
machine-learning based classification. Various candidate fea-
ture extraction/selection algorithms were discussed, in order to
reduce the predictive variables from the 48 extracted through
image processing techniques. The resulted datasets were fed
in 12 classifiers.
One of the lessons learned is that greedy search, best
first search with backward search and principal component
analysis should be avoided for this problem, as their induced
performance is constantly poor across all classifiers. Although
it is hard to extrapolate to other image classification problems,
one should not blindly use them.

972
 Another remark is that it is worth investigating the potential
benefits of feature selection. The original 48 numeric attributes
allowed statistically better performance in 4 cases, but also
performed poorer in 3 cases. Some feature selection techniques
– evolutionary and genetic algorithms, reranking search, best-
first with forward search and linear forward selection – repeat-
edly allowed the classifiers to outperform the baseline model.
Again, it is hard to extrapolate from a single case study, but
it is reasonable that one might expect attaining better results
when one filters out some predictive variables.
Finally, we remarked that the two classifiers with the
largest number of good results rely on local models, building
predictions based on data that cluster together. This behavior
deserves further investigation, as the literature offers plenty
of such cluster–based associators. It might also be worth
comparing these results with the ones produced by deep
learning architectures: they have the potential to automatically
build a set of features on top of which a classifier might lead
to a better performance. To our best knowledge, such a study
for small datasets of images is not yet performed.
R EFERENCES
[1] T. M. Mitchell, “The need for biases in learning generalizations,” Rutgers
University, New Brunswick, NJ, Tech. Rep., 1980.
[2] Y. Bengio, “Learning deep architectures for AI,” Foundations and Trends
in Machine Learning, vol. 2, no. 1, pp. 1–127, 2009, also published as
a book. Now Publishers, 2009.
[3] Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A
review and new perspectives,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 35, no. 8, pp. 1798–1828, 2013.
[4] I. Arel, D. C. Rose, and T. P. Karnowski, “Research Frontier: Deep
Machine Learning–a New Frontier in Artificial Intelligence Research,”
Comp. Intell. Mag., vol. 5, no. 4, pp. 13–18, Nov. 2010.
[5] A. Graves, Supervised Sequence Labelling with Recurrent Neural Net-
works, ser. Studies in Computational Intelligence. Springer, 2012, vol.
385.
[6] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” in Intelligent Signal Processing.
IEEE Press, 2001, pp. 306–351.
[7] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classifica-
tion with deep convolutional neural networks,” in Advances in Neural
Information Processing Systems 25, P. Bartlett, F. Pereira, C. Burges,
L. Bottou, and K. Weinberger, Eds., 2012, pp. 1106–1114.
[8] K. Kavukcuoglu, P. Sermanet, Y.-L. Boureau, K. Gregor, M. Mathieu,
and Y. L. Cun, “Learning convolutional feature hierarchies for visual
recognition,” in Advances in Neural Information Processing Systems 23,
J. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. Zemel, and A. Culotta,
Eds., 2010, pp. 1090–1098.
[9] D. C. Ciresan, U. Meier, L. M. Gambardella, and J. Schmidhuber,
“Convolutional neural network committees for handwritten character
classification,” in ICDAR, 2011, pp. 1135–1139.
[10] ——, “Deep, big, simple neural nets for handwritten digit recognition,”
Neural Computation, vol. 22, no. 12, pp. 3207–3220, 2010.
[11] M. Langford, G. Taylor, and J. Flenley, “Computerized identification
of pollen grains by texture analysis,” Review of Palaeobotany and
Palynology, vol. 1, no. 4, pp. 197–203, Nov. 2010.
[12] W. Treloar, Digital Image Processing Techniques and Their Application
to the Automation of Palynology, Ph.D. thesis. University of Hull, 1992.
[13] M. Rodrguez-Damin, E. Cernadas, A. Formella, and P. de S-Otero,
“Pollen classification using brightness-based and shape-based descrip-
tors.” in ICPR (2), 2004, pp. 212–215.
[14] Y. Zhang, R. Wang, and P. Hunter, “Airborne pollen texture discrim-
ination using wavelet transforms in combination with cooccurrence
matrices.” Int. J. of Intelligent Systems Technologies and Applications,
vol. 1, no. 1/2, pp. 143–156, 2005.
973
[15] R. Dell’anna, P. Lazzeri, M. Frisanco, F. Monti, C. F. Malvezzi,
E. Gottardini, and M. Bersani, “Pollen discrimination and classification
by fourier transform infrared (ft-ir) microspectroscopy and machine
learning,” Anal Bioanal Chem, vol. 394, no. 5, pp. 1443–1452, 2009.
[16] A. Boucher and M. Thonnat, “Object recognition from 3d blurred
images.” in ICPR (1), 2002, pp. 800–803.
[17] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 3rd ed.
Pearson Education International, 2008.
[18] M. del Pozo-Baños, J. R. Ticay-Rivas, J. Cabrera-Falcón, J. Arroyo,
C. M. Travieso-González, L. Sánchez-Chavez, S. T. Pérez, J. B. Alonso,
and M. Ramı́rez-Bogantes, “Image Processing For Pollen Classification,”
InTech, 2012.
[19] F. S. Perkins, A. Walker, and E. Wolfart, Hypermedia Image Processing
Reference. Wiley, 1997.
[20] I. Guyon and A. Elisseeff, “An introduction to variable and feature
selection,” J. Mach. Learn. Res., vol. 3, pp. 1157–1182, Mar. 2003.
[21] I. Guyon, S. Gunn, M. Nikravesh, and L. A. Zadeh, Feature Extraction:
Foundations and Applications (Studies in Fuzziness and Soft Comput-
ing). Secaucus, NJ, USA: Springer-Verlag New York, Inc., 2006.
[22] H. Liu and H. Motoda, Computational Methods of Feature Selection.
Chapman & Hall/CRC, 2007.
[23] Y. Saeys, I. n. Inza, and P. Larrañaga, “A review of feature selection
techniques in bioinformatics,” Bioinformatics, vol. 23, no. 19, pp. 2507–
2517, Sep. 2007.
[24] J. Fan and J. Lv, “A selective overview of variable selection in high
dimensional feature space,” Statistica Sinica, vol. 20, no. 1, pp. 101–
148, 2010.
[25] T. J. Hastie, R. J. Tibshirani, and J. H. Friedman, The elements of
statistical learning : data mining, inference, and prediction, ser. Springer
series in statistics. New York: Springer, 2009.
[26] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and
I. H. Witten, “The WEKA data mining software: an update,” SIGKDD
Explorations Newsletter, vol. 11, no. 1, pp. 10–18, 2009.
[27] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Ma-
chine Learning, 1st ed. Boston, MA, USA: Addison-Wesley Longman
Publishing Co., Inc., 1989.
[28] M. Gütlein, E. Frank, M. Hall, and A. Karwath, “Large-scale attribute
selection using wrappers,” in Proc. IEEE Symposium on Computational
Intelligence and Data Mining, 2009, pp. 332–339.
[29] A. Moraglio, C. D. Chio, and R. Poli, “Geometric particle swarm
optimisation.” in EuroGP, ser. Lecture Notes in Computer Science,
M. Ebner, M. O’Neill, A. Ekrt, L. Vanneschi, and A. Esparcia-Alczar,
Eds., vol. 4445. Springer, 2007, pp. 125–136.
[30] P. Bermejo, L. de la Ossa, J. A. Gámez, and J. M. Puerta, “Fast wrapper
feature subset selection in high-dimensional datasets by means of filter
re-ranking,” Know.-Based Syst., vol. 25, no. 1, pp. 35–44, Feb. 2012.
[31] A.-R. Hedar, J. Wang, and M. Fukushima, “Tabu search for attribute
reduction in rough set theory.” Soft Comput., vol. 12, no. 9, pp. 909–
918, 2008.
[32] R. Caruana and A. Niculescu-Mizil, “An empirical comparison of
supervised learning algorithms,” in Proceedings of the 23rd International
Conference on Machine Learning, ser. ICML ’06. New York, NY, USA:
ACM, 2006, pp. 161–168.
[33] E. Frank and I. H. Witten, “Generating accurate rule sets without global
optimization,” in Proceedings of the Fifteenth International Conference
on Machine Learning, ser. ICML ’98. San Francisco, CA, USA:
Morgan Kaufmann Publishers Inc., 1998, pp. 144–151.
[34] J. Gama, “Functional trees,” Mach. Learn., vol. 55, no. 3, pp. 219–250,
Jun. 2004.
[35] N. Landwehr, M. Hall, and E. Frank, “Logistic model trees,” Mach.
Learn., vol. 59, no. 1-2, pp. 161–205, May 2005.
[36] H. Shi, “Best-first decision tree learning,” Master’s thesis, University of
Waikato, Hamilton, NZ, 2007, cOMP594.
[37] G. Hulten, L. Spencer, and P. Domingos, “Mining time-changing data
streams,” in Proceedings of the Seventh ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, ser. KDD ’01.
New York, NY, USA: ACM, 2001, pp. 97–106.
[38] J. R. Quinlan, C4.5: Programs for Machine Learning. San Francisco,
CA, USA: Morgan Kaufmann Publishers Inc., 1993.
[39] L. Breiman, “Random forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32,
Oct. 2001.
[40] C. Nadeau and Y. Bengio, “Inference for the generalization error,” Mach.
Learn., vol. 52, no. 3, pp. 239–281, Sep. 2003.