AUTOMATIC HOVERFLY SPECIES DISCRIMINATION

Branko Brklja

, Marko Pani

, Dubravko

Culibrk

, Vladimir Crnojevi

, Jelena A

canski

and Ante Vuji

Department of Power, Electronics and Communication Engineering, Faculty of Technical Sciences, University of Novi Sad

Trg Dositeja Obradovi

ca 6, 21000, Novi Sad, Serbia

Department of Biology and Ecology, Faculty of Sciences, University of Novi Sad

Trg Dositeja Obradovi

ca 2, 21000, Novi Sad, Serbia

Keywords:

Species discrimination, Wing venation, Junctions detection, Support vector machine, HOG, LBP.

Abstract:

An novel approach to automatic hoverﬂy species discrimination based on detection and extraction of vein

junctions in wing venation patterns of insects is presented in the paper. The dataset used in our experiments

consists of high resolution microscopic wing images of several hoverﬂy species collected over a relatively long

period of time at different geographic locations. Junctions are detected using histograms of oriented gradients

and local binary patterns features. The features are used to train an SVM classiﬁer to detect junctions in wing

images. Once the junctions are identiﬁed they are used to extract simple statistics concerning the distances

of these points from the centroid. Such simple features can be used to achieve automatic discrimination of

four selected hoverﬂy species, using a Multi Layer Perceptron (MLP) neural network classiﬁer. The proposed

approach achieves classiﬁcation accuracy of environ 71%.

1 INTRODUCTION

Classiﬁcation, measurement and monitoring of in-

sects form an important part of many biodiversity and

evolutionary scientiﬁc studies (Houle et al., 2003),

(Arbuckle et al., 2001), (Larios et al., 2008). Their

aim is usually to identify presence and variation of

some characteristic insect or its properties that could

be used as a starting point for further analysis. The

technical problem that researchers are facing is a very

large number of species, their variety, and a shortage

of available experts that are able to categorize and ex-

amine specimens in the ﬁeld. Due to these circum-

stances, there is a constant need for automation and

speed up of this time consuming process. Applica-

tion of computer vision and its methods provides ac-

curate and relatively inexpensive solutions when ap-

plicable, as it is in the case of different ﬂying in-

sects (Houle et al., 2003), (MacLeod, 2007), (Ar-

buckle et al., 2001), (Zhou et al., 1985). Some of

these insects are pollinators that play a great role in

nature and are of particular interest for scientists as

important indicator species. Their wings are one of

the most frequent discriminating characteristics con-

sidered (MacLeod, 2007) and can be used standalone

or as a key characteristic for insect classiﬁcation (Ar-

buckle et al., 2001). Unlike some other body parts,

wings are also particularly suitable for automatic pro-

cessing (Toﬁlski, 2008). The process can be aimed at

species identiﬁcation and classiﬁcation, or form the

basis of further morphometric analyses once the clas-

siﬁcation to speciﬁc taxonomy is done.

Discriminative information that allows ﬂying in-

sects classiﬁcation may be contained in wing shape

(Rohlf and Archie, 1984), but in most cases it is con-

tained in the relative positions of vein junctions inside

wing that mostly deﬁne unique wing venation pat-

terns (Houle et al., 2003), (MacLeod, 2007), (Toﬁlski,

2008), (Arbuckle et al., 2001), (Zhou et al., 1985).

Wing venation patterns are the result of speciﬁc evo-

lutionary adaptations over a long period of time and

are inﬂuenced by many different factors (Thompson,

1945). As such, they are relatively stable and can suc-

cessfully describe and represent small differences be-

tween very similar species and taxons, what is not al-

ways possible using only wing shape. Another use-

ful property is that they are not affected signiﬁcantly

by the current living conditions, present in some spe-

ciﬁc natural environment, when compared to some

other wing properties such as colour or pigmentation.

This makes them a good choice for reliable and robust

species discrimination and measurement. The advan-

tage of using venation patterns is also that patterns of

previously collected wing specimens do not change

108

Brklja

c B., Pani

c M., Ä ˛Eulibrk D., Crnojevi

c V., A

canski J. and Vuji

c A. (2012).

AUTOMATIC HOVERFLY SPECIES DISCRIMINATION.

In Proceedings of the 1st International Conference on Pattern Recognition Applications and Methods, pages 108-115

DOI: 10.5220/0003756601080115

 SciTePress

with the pass of time, as some other wing features, so

they are suitable for later, off-ﬁeld analyses. Discrim-

ination of species in the past was based on descrip-

tive methods that proved to be insufﬁcient and were

replaced by morphometric methods (Toﬁlski, 2008).

These methods rely on geometric measures like an-

gles and distances in the case of standard morphom-

etry or coordinates of key points called landmarks,

that could be also used for computing angles and dis-

tances, in the case of more recent geometric morpho-

metrics. In wing-based discrimination each landmark

point represents a unique vein junction that is previ-

ously determined. Manually determined landmarks

require skilled operator and are prone to errors, so au-

tomatic detection of landmark points is always pre-

ferred.

Most state-of-the-art systems for insect classiﬁca-

tion contain, in addition to equipment for specimens

handling, components for image acquisition and anal-

ysis that enable extraction of speciﬁc discriminative

information to base specimen classiﬁcation on. Main

differences between the systems relate to the type of

information they look for and the way it is obtained

from image. Some are designed to perform recog-

nition tasks in uncontrolled environments with vari-

ability in position and orientation of objects (Larios

et al., 2008), and other work under controlled condi-

tions (Toﬁlski, 2008), (Arbuckle et al., 2001). Unfor-

tunately they are usually not general in their applica-

tion.

Methods for automatic detection of vein junctions

in wing venation of insects are usually based on sim-

ilar computer vision techniques. They generally con-

sist of several preprocessing steps that include image

registration, wing segmentation, noise removal and

contrast enhancement. In order to extract lines that

deﬁne wing venation pattern, edge detection, adap-

tive thresholding, morphological ﬁltering, skeleton

extraction, pruning and interpolation are often applied

in next stage. Thus, the landmark points correspond-

ing to vein junctions are found (Houle et al., 2003),

(MacLeod, 2007) or a polynomial model of whole

venation pattern is made on the base of line junctions

and intersections (Houle et al., 2003), (Arbuckle et al.,

2001), (Zhou et al., 1985). In both cases, the main

prerequisite is to obtain an image that contains only

wing outline and wing venation skeleton. That may

be easier to achieve if the light source is precisely

aligned during the image acquisition phase so that it

produces uniform background (MacLeod, 2007), or

when it is allowed to use additional colour informa-

tion as in the case of leaf venation patterns (Zheng

and Wang, 2009), but it is not always the case. Some

of the possible reasons are noisy and damaged images

due to dust, pigmentation, different wing sizes, image

acquisition or bad specimen handling. Another obsta-

cle is that at each processing stage there are numerous

choices and different solutions that are in most cases

problem-dependent. As a result, currently available

systems and algorithms are very specialized and con-

tain different problem speciﬁc adaptations.

The goal of the research presented here is to de-

velop an automated ﬂying insects identiﬁcation sys-

tem based on wing venation patterns, primarily in-

tended for hoverﬂies, family Syrphidae. The paper

presents an approach to hoverﬂy species discrimina-

tion based on a novel method for automatic detec-

tion of landmark points in wing venation of insects.

Instead of using problem-dependent algorithms for

wing venation skeleton extraction, we propose the use

of a machine learning algorithm trained on a vein

junctions dataset extracted by human-experts from

real-world images.

The rest of the paper is organized as follows. Sec-

tion 2 provides an overview of the dataset and the

landmark-points detection method used. The pro-

posed hoverﬂy-species-discrimination methodology

is presented in Section 3. Evaluation results are given

in Section 4 and conclusions are drawn in Section 5.

2 LANDMARK POINTS

DETECTION

The proposed method for landmark point (vein junc-

tions) detection consists of computing speciﬁc, win-

dow based features (Ojala et al., 1996), (Dalal and

Triggs, 2005), (Wang et al., 2009), which describe

presence of textures and edges in window, and sub-

sequent classiﬁcation of these windows as junctions

(positives) or not-junctions (negatives) using detector

obtained by some supervised machine learning tech-

nique.

2.1 Wing Images Dataset

The set of wing images used in the study presented

consists of high-resolution microscopic wing images

of several hoverﬂy species. There are 868 wing im-

ages of eleven hoverﬂy species from two different

genera: Chrysotoxum and Melanostroma, Table 1.

Table 1: Number of wing images per class.

Chrysotoxum Melanostroma

Festivum Vernale other Mellinum Scalare other

248 154 22 267 105 72

AUTOMATIC HOVERFLY SPECIES DISCRIMINATION

109

Wings were collected from many different geo-

graphic locations during a relatively long period of

time of more than two decades. Wing images are

obtained from wings mounted in glass microscopic

slides using a microscopic device with a camera res-

olution 2880×1550 pixels and stored in TIFF format.

Each image is uniquely numbered and sorted to the

group it belongs. Association of each wing with a

particular species is based on classiﬁcation of the in-

sect at the time it was collected and before wings were

detached. This classiﬁcation was done after exami-

nation by an experienced expert. The images them-

selves were acquired later by biologist under rela-

tively uncontrolled conditions of nonuniform back-

ground light and variable scene conﬁguration without

previous camera calibration. In that sense, obtained

images are not particularly suitable for exact measure-

ments.

Other shortcomings that occur in the dataset, are

result of variable wing specimens quality, damaged

or bad mounted wings, existence of artifacts, variable

wing positions and dust. In order to overcome these

limitations and make these images amenable to auto-

matic hoverﬂy species discrimination, they were pre-

processed. The preprocessing consisted of image ro-

tation to a uniﬁed horizontal position, wing cropping

and scaling. Cropping eliminates unnecessary back-

ground containing artifacts. After the calculation of

mean width and height of all cropped images, they

were interpolated to the same size of 1680×672 pix-

els using bicubic interpolation. Wing images obtained

on this way form the ﬁnal wing-image dataset used

for sliding-window detector training, its performance

evaluation and subsequent hoverﬂy species discrim-

ination. Number of images per species is not uni-

form, Table 1, so only four species with signiﬁcant

number of images are selected for later discrimina-

tion based on detected landmark points, Fig.1. These

four species relate to 774 images and belong to both

genera of the Syrphidae family.

2.2 Training/Test Set

In order to analyze applicability and efﬁciency of the

proposed methodology when it comes to the prob-

lem of landmark-point detection, special vein junc-

tions training/test set was created from collected wing

images. It consists of characteristic wing regions

(patches) that correspond to vein junctions and ran-

domly selected patches, negatives without vein junc-

tions. From each wing image 18 uniquely numbered

positive patches, shown on Fig. 1, were manually ex-

tracted and saved using specially created application.

In the case of severely damaged wings there were

Figure 1: Selected hoverﬂy species from two diff. gen-

era (from top to bottom): Chrysotoxum Festivum, Chryso-

toxum Vernale, Melanostroma Mellinum, Melanostroma

Scalare.

damaged or missing landmarks that were not selected.

As a result, the total number of positives is slightly

smaller than expected and the training/test set with

15590 positives and 22466 manually selected nega-

tives was created. During detector evaluation some of

the selected landmarks (landmarks numbered 0 and

1 on Fig. 1) prove to be not descriptive enough to

properly and reliably describe wing vein junctions, al-

though they were ﬁrstly marked as a landmarks, so

they were discarded from further analysis. Reason is

their grater variability due to speciﬁc position, which

in combination with relatively small dataset makes

their detection and even proper selection during train-

ing phase harder. After discarding these landmarks

the ﬁnal training/test set containing 13868 positives

from all available hoverﬂy wing images was obtained.

The created set was then used for detailed study of

various implementation choices on detector perfor-

mance.

ICPRAM 2012 - International Conference on Pattern Recognition Applications and Methods

110

2.3 Landmark-point Detector

Discriminative descriptors of vein junctions that are

used in the proposed landmark-point detector are

HOG (Histogram of Oriented Gradients) and LBP

(Local Binary Pattern). They were ﬁrst proposed in

(Dalal and Triggs, 2005) and (Ojala et al., 1996). In

order to determine and compare their performance

and evalutate the impact of different sets of pertinent

parameters, they were considered separately and com-

bined as described in detail in (Wang et al., 2009).

When it comes to HOG, a feature vector consist-

ing of a number of discrete histograms of image gradi-

ent orientation (Dalal and Triggs, 2005) is employed.

Discrete histograms are computed over small rectan-

gular spatial regions, called cells, that are obtained

by subdivision of the main window. The ﬁrst step

in the histogram computation is gradient discretiza-

tion, done for each pixel in two closest allowed bins.

Before computing discrete histograms for each cell,

2-D CTI ﬁltering described in (Wang et al., 2009) is

applied. As suggested in (Dalal and Triggs, 2005),

before the construction of the ﬁnal feature vector, his-

togram values are locally normalized by accumulating

histograms over somewhat larger overlapping spatial

regions, called blocks, using the L

norm. These val-

ues, representing normalized values of several spa-

tially adjacent discrete histograms, are than serially

written, block by block, to form the ﬁnal HOG fea-

ture vector.

Vector length and the dimensionality of corre-

sponding feature space depend on the choice of pa-

rameters that deﬁne window, cell and block size,

block overlapping and a number of allowed discrete

histogram values (orientation bins). We used nine

bins evenly spaced over 0

◦

-180

◦

, a 64 × 64 detec-

tion window, blocks containing 2 × 2 cells and block

overlapping width one cell wide. In order to measure

detector performance different cell sizes (8, 16 and

32 pixels) were used. As a result, depending on cell

size, possible dimensions of HOG feature vectors are:

1764 (hog8), 324 (hog16) and 36 (hog32). Fig. 2.

shows the sample HOG feature of a junction (land-

mark point).

An LBP feature describes local structure of the

pixel values in a grayscale image. It is computed for

every pixel by comparing its value P

with the val-

ues of all surrounding pixels P

, using a predeﬁned

rule. The rule deﬁnes the neighbourhood, its geom-

etry and the way of comparing pixel values. Results

of comparisons are written as binary code that repre-

sents a particular pattern in the local neighbourhood

of the pixel considered. As a measure for comparing

pixel values we used quantized difference between

(a) hog8

(b) hog16 (c) hog32

Figure 2: Illustration of HOG feature extraction on the ex-

ample of vein junction from training/test set using different

cell sizes (8, 16, and 32 pixels respectively).

surrounding and central pixel values. If we denote

quantized pixel difference value with a(·), a binary

pattern LBP

n,r

that describes texture is:

LBP

n,r

n−1

∑

i=0

a(P

− P

, a(x) =



1, x ≥ 0

0, x < 0

(1)

where n denotes number of neighbouring pixels in

the radius r that are compared with the central pixel.

Rectangular neighbourhood geometry with surround-

ing pixels at unit distance, 16 × 16 cells and the same

window size as in the case of HOG were used.

Under LBP feature, obtained for each analyzed

window, we assume feature vector consisting of val-

ues that represent several discrete histograms of LBP

codes. Histograms are computed separately for each

cell in a window, by counting LBP codes previously

calculated for all cell pixels. Histogram values are

normalized using L

norm without any overlap be-

tween the cells. Of particular interest are only uni-

form LBP codes (Wang et al., 2009) that are repre-

sented as separate values in the histogram. Uniform

AUTOMATIC HOVERFLY SPECIES DISCRIMINATION

111

Figure 3: Detector feature performance comparison using

linear SVM.

codes LBP

n,r

are those having no more than u 0-1 or

1-0 transitions. All other codes are grouped in one

class, corresponding to the last histogram value. For

the purpose of vein junctions extraction we used 58

uniform LBP

8,1

codes. As a result dimensionality of

the LBP feature space is 944.

Combined feature vectors are formed by append-

ing LBP feature vector at the end of HOG feature vec-

tor as described in (Wang et al., 2009). Both HOG and

LBP feature vectors were used separately and in all

combinations in order to measure their window based

performance on the training/test set using the same

classiﬁer. Performance comparison was made using

linear Support Vector Machine (SVM) classiﬁer that

has good generalization properties and ability to cope

with small number of samples in the case of high fea-

ture space dimensionality (Vapnik, 2000).

Feature extraction was implemented in C++ using

OpenCV library (Bradski and Kaehler, 2007). De-

tector performance testing was done in the machine-

learning package Weka (Hall et al., 2009) using Lib-

SVM library (Chang and Lin, 2001), which contains

an implementation of an SVM classiﬁer. In all cases

classiﬁer performance was measured using 10 fold

cross-validation. Cross-validated window level re-

sults in terms of true positives and false positives rates

on training/test set are shown on Fig. 3.

Usage of HOG and LBP features as descriptors of

vein junctions shows acceptable results with miss rate

smaller than 3% in most cases. When used separately

LBP feature gives better result than HOG feature.

HOG feature with 32 pixels cell size is too coarse to

properly describe vein junction in the middle of the

window, because in this case window contains only 4

cells. On the other hand the smallest cell size of 8 pix-

els gave best result among HOG features. As can be

seen from the Fig. 3. combined features are best in the

sense of performance but are more memory and time

demanding during the training phase. Nevertheless,

shown results were motivation for the construction of

vein junctions sliding window detector.

As a result combined HOG-LBP feature with the

cell size of 16 pixels was selected as the best choice

for automatic hoverﬂy species discrimination based

on sliding window landmarks detection.

3 SPECIES DISCRIMINATION

Automatic hoverﬂy species discrimination was lim-

ited only to four selected hoverﬂy species from the

wing-image dataset with sufﬁcient number of in-

stances. The discrimaination is based on the output

of the automatic detection of vein junctions described

in Section 2. Vein-junction detection is done using

a sliding window to search through the image and

the detector described in Section 2. For better per-

formance nonlinear SVM classiﬁer implemented in

(Bradski and Kaehler, 2007) is used. Its optimal pa-

rameters are determined through exponential parame-

ter grid search using 10 fold cross-validation on whole

training/test set. Once the optimal values of the pa-

rameters are determined, the train/test dataset is used

to train the ﬁnal detector. The constructed detector

goes through the wing image and returns discrete re-

sponses indicating whether a vein junction is present

or not in current window. The same size of the slid-

ing window step is used for both image dimensions.

In the case of a detection, window center coordinates

corresponding to possible vein junction are stored to-

gether with classiﬁers soft response value. This value

describes how far from separating hyperplane deﬁned

by support vectors, the current feature vector is, or

how trustworthy the detectors decision is. This soft

information is later used to improve the precision of

landmark detections.

Due to multiple detections of the same vein joint,

additional postprocessing of obtained detections is

needed, once the detector ﬁnishes searching through

the image. The postprocessig consists of point clus-

terization, Fig. 4, and subsequent computation of

each clusters centroid using previously obtained de-

tectors soft response values. Clusterization is based

on the iterative algorithm that searches through the

remaining detections that have not yet been associ-

ated with some existing cluster until all detections are

assigned to some cluster. It uses a distance criterion

based on the sliding window step size and initializes

clusters with existing unassociated detections. Once

the clusterization is completed, the centroid of each

cluster is determined as weighted average of all de-

tections inside the cluster. In order to use the soft re-

sponse values as weighting factors, these values are

normalized on the level of cluster using L

norm so

that they correspond to probabilities of true vein junc-

ICPRAM 2012 - International Conference on Pattern Recognition Applications and Methods

112

Figure 4: Automatic detections of vein junctions in wing images after clusterization, clusters are shown with different colours.

tion detection. Obtained cluster centroids represent

vein junctions that have been found in an image by

sliding window detector.

However there is no guarantee that these au-

tomatic detections contain all expected landmark

points, or that there are no false detections. The rea-

sons for this can be damaged wings, presence of arti-

facts, etc. Consequently, ﬁxed length feature vectors,

which would be based on obtained automatic detec-

tions, are not an appropriate choice for image classiﬁ-

cation. Therefore, we propose an approach that is not

sensitive to number of detected landmark points. A

AUTOMATIC HOVERFLY SPECIES DISCRIMINATION

113

Table 2: True positives (TP), false positives (FP) rate, precision and recall.

Chrysotoxum Melanostroma Weighted

Festivum(a) Vernale(b) Mellinum(c) Scalare(d) average

TP Rate 0.774 0.383 0.888 0.543 0.704

FP Rate 0.181 0.103 0.093 0.034 0.115

Precision 0.669 0.480 0.835 0.713 0.694

Recall 0.774 0.383 0.888 0.543 0.704

Table 3: Classiﬁcation results, confusion matrix for four

selected hoverﬂy species.

Chrysotoxum Melanostroma Total

Festivum Vernale Mellinum Scalare samples

a b c d

a 192 54 1 1 248

b 90 59 4 1 154

c 2 7 237 21 267

d 3 3 42 57 105

feature vector consisting of landmark points centroid,

average Euclidean distance of landmark points from

centroid, median of these distances and their standard

deviation around average is constructed. A common

property of these measures is that they, as descriptive

statistics, do not depend signiﬁcantly on the number

of landmark points used for their computation and are

also rotational invariant. Under the assumption that

they are discriminative enough to distinguish differ-

ent hoverﬂy species and do not change signiﬁcantly

inside the same species, they are used as part of fea-

ture vector that describes particular wing image. As

a result dimensionality of used feature space is only

ﬁve. In order to eliminate possible false detections

only landmark points that originate from clusters with

more than two detections are used. Automatic dis-

crimination of four selected hoverﬂy species is then

made using Multi Layer Perceptron (MLP) neural net-

work classiﬁer with single hidden layer consisting of

4 neurons implemented in (Hall et al., 2009), Table 3.

4 RESULTS

The performance of automatic landmarks detection

using different sliding window step sizes was ana-

lyzed. Step sizes of 8, 16 and 32 pixels were used and

different degrees of landmark detections per image

were achieved. Using the sliding window with largest

step size is signiﬁcantly faster than the alternatives,

but with the smallest number of detected landmarks

per image and most imprecise due to the absence of

multiple detections. The highest detection accuracy

was achieved using the smallest step size, so this de-

tector was selected to serve as basis for species classi-

ﬁcation using MLP. Classiﬁer results, obtained using

10 fold cross-validation, are given in Table 2 and 3.

Classiﬁcation accuracy of 70.4% was achieved.

These results conﬁrm the applicability of pro-

posed approach in the sense that used features, based

on obtained automatic detections, enable very high

discrimination between two genera inside the same

family, Table 3. Usage of same species dicrimina-

tion approach with ground-truth landmark data in-

stead of automatic landmark detections gave accuracy

of 74.6% correctly classiﬁed instances.

5 CONCLUSIONS

Systems for automatic insects classiﬁcation are usu-

ally designed for ﬁeld use. Therefore it is desirable

that they are robust and as general as possible. At

present time image based systems are considered as

the preferred choice comparing to some other alter-

natives, like DNA analysis. An image processing ap-

proach to hoverﬂy species discrimination presented in

this paper shows good results on the collected wing-

image dataset. Its advantage is that it is based on

proposed robust method for the detection of land-

mark points in wing venation of insects that can cope

with different image imperfections, Fig. 4. Simple

rotation-invariant features chosen for later wing clas-

siﬁcation are one possible solution for the problem

of unpredictable number of automatic detections and

proved to be discriminative enough to distinguish cor-

rectly between two different hoverﬂy genera and to

a lesser extent, the species that comprise them. As

such, it could be used as a ﬁrst step in a construction

of a more complex cascade classiﬁer that would in-

corporate extra information from the image and from

the obtained landmark points to achieve more precise

classiﬁcation on the species level.

REFERENCES

Arbuckle, T., Schr

oder, S., Steinhage, V., and Wittmann, D.

(2001). Biodiversity informatics in action: identiﬁ-

cation and monitoring of bee species using ABIS. In

ICPRAM 2012 - International Conference on Pattern Recognition Applications and Methods

114

Proc. 15th Int. Symp. Informatics for Environmental

Protection, pages 10–12.

Bradski, G. and Kaehler, A. (2007). Learning OpenCV.

O’Reilly Media, Sebastopol CA.

Chang, C. and Lin, C. (2001). Libsvm: a library for sup-

port vector machines. ACM Transactions on Intelli-

gent Systems and Technology), 2(3):27.

Dalal, N. and Triggs, B. (2005). Histograms of oriented

gradients for human detection. In CVPR 2005, IEEE

Computer Society Conference on Computer Vision

and Pattern Recognition, volume 1, pages 886–893.

Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reute-

mann, P., and Witten, I. (2009). The weka data min-

ing software: an update. ACM SIGKDD Explorations

Newsletter, 11(1):10–18.

Houle, D., Mezey, J., Galpern, P., and Carter, A. (2003).

Automated measurement of Drosophila wings. BMC

Evolutionary Biology, 3(1):25.

Larios, N., Deng, H., Zhang, W., et al. (2008). Automated

insect identiﬁcation through concatenated histograms

of local appearance features: feature vector generation

and region detection for deformable objects. Machine

Vision and Applications, 19(2).

MacLeod, N., editor (2007). Automated taxon identiﬁcation

in systematics: theory, approaches and applications,

pages 289–298. CRC Press, Boca Raton.

Ojala, T., Pietik

ainen, M., and Harwood, D. (1996). A com-

parative study of texture measures with classiﬁcation

based on featured distributions. Pattern Recognition,

29(1):51–59.

Rohlf, F. and Archie, J. (1984). A comparison of

Fourier methods for the description of wing shape in

mosquitoes (Diptera: Culicidae). Systematic Biology,

33(3):302.

Thompson, W. (1945). On growth and form. Cambridge

University Press, Cambridge.

Toﬁlski, A. (2008). Using geometric morphometrics and

standard morphometry to discriminate three honeybee

subspecies. Apidologie, 39(5):558–563.

Vapnik, V. (2000). The nature of statistical learning theory.

Springer Verlag.

Wang, X., Han, T., and Yan, S. (2009). An HOG-LBP hu-

man detector with partial occlusion handling. In ICCV

2009, 12th International Conference on Computer Vi-

sion, pages 32–39. IEEE.

Zheng, X. and Wang, X. (2009). Fast Leaf Vein Extrac-

tion Using Hue and Intensity Information. In ICIECS

2009,International Conference on Information Engi-

neering and Computer Science, pages 1–4. IEEE.

Zhou, Y., Ling, L., and Rohlf, F. (1985). Automatic descrip-

tion of the venation of mosquito wings from digitized

images. Systematic Biology, 34(3):346.

AUTOMATIC HOVERFLY SPECIES DISCRIMINATION

115