classification of dates varieties and effect of motion blurring

7/29/2019 Classification of Dates Varieties and Effect of Motion Blurring

1/6

O R I G I N A L P A P E R

Classification of dates varieties and effect of motion blurringon standardized moment features

Gabriel Thomas A. Manickavasagan

R. Al-Yahyai

Received: 30 July 2012 / Accepted: 11 October 2012

Springer Science+Business Media New York 2012

Abstract Computer vision technology has been used as

a successful non-destructive quality assessment tool forvarious food products. In general, several features are

extracted from the images of interest, and used for the

classification models. Furthermore, in most of the studies,

static images have been used in the calibration and evalu-

ation models. Classification models with a reduced number

of features, and a mechanism to test the capability of the

algorithm for moving objects by means of simulating the

blurring effect on the static images would be beneficial to

determine the performance of the system in real-time

quality monitoring in industries. Using three date varieties

as model food, motion was simulated for the dates images

and a successful neural network classifier was designed

with only three statistical features (mean, standard devia-

tion, and skewness). The reduced number of features and

simplicity of the classifier yielded a solution that can be

potentially implemented in hardware fast enough so that to

consider the case of classification of the dates in a conveyor

belt. To test the solution under such conditions, a blurring

degradation function was used to verify that the classifierwould work. The effects that motion blurring causes to

these statistical moments in a general sense were examined

using random numbers drawn from the distribution in the

Pearson system. Because motion blurring showed a ten-

dency to change the distribution to a Gaussian density, the

same features and classifier yielded similar results despite

of motion.

Keywords Classification Neural network Bayes classifier Pearson random numbers Statistical

moment Image motion

Introduction

In computer vision (CV) technology, objects are imaged

and analyzed to characterize their quality. This technique

has great potential to be used as a non-destructive and

objective quality measurement method. It is a reliable

technique for the measurement of various quality attributes

of agricultural and food products [14]. Attribute charac-

terization using images taken while the objects are at static

could be implemented at quality control laboratories.

However, the efficiency of the developed classification

models for the objects moving on a conveyor at various

speeds would be highly beneficial to determine their ability

for online quality monitoring in the real-time production

facilities in food industries. Therefore the objective of this

study was to determine the classification effects caused by

simulated motion of date varieties with three grayscale

features at static and motion blurred conditions. Motion

was simulated via low pass filtering which system response

is usually calculated in order to eliminate motion bluring

G. Thomas (&)

Department Electrical and Computer Engineering, Faculty of

Engineering, University of Manitoba, E3-555 Engineering and

Information Technology Complex, Winnipeg, MB R3T5V6,

Canadae-mail: [email protected];

[email protected]

A. Manickavasagan

Department of Soils, Water and Agricultural Engineering,

College of Agricultural and Marine Sciences, Sultan Qaboos

University, P.O. Box 34, 123 Al-Khoud, Sultanate of Oman

R. Al-Yahyai

Department of Crop Science, College of Agricultural and Marine

Sciences, Sultan Qaboos University, P.O. Box 34,

123 Al-Khoud, Sultanate of Oman

123

Food Measure

DOI 10.1007/s11694-012-9129-9


2/6

using deconvolution techniques. Different methods propose

alternative models for the low pass impulse response of

the system and the expression that corresponds to linear

motion in one direction as defined in [12] is used here; a

conveyor belt motion scenario reduces the model to this

case only. Dates are used as model images and this tech-

nique can be used as such in other products.

Date is an important commodity in Oman, and around50 % of the cultivable lands are under date palm vegeta-

tion. Although annual production of dates in Oman is

255,891 Mt, only 9,000 Mt (2.53.5 % of production) is

exported due to various reasons [5]. Quality assurance has

always been a major problem for Omani dates to compete

in international markets [6]. Varietal purity, color, unifor-

mity of size, and absence of defects are some of the

important quality parameters for dates in domestic and

international market. Generally, manual grading of dates is

followed in handling and processing facilities to identify

date varieties. This method has many constraints such as

subjectivity, influence of mental stress, influence of envi-ronment, efficiency of individuals at various times of the

shift, and so on. An automated variety identification

method using CV technique would be beneficial to the date

industries in Oman which would correspond to a first step

to assess the quality by avoiding contamination from dif-

ferent varieties.

Materials and methods

Image acquisition

Three date varieties namely Khalas, Fard, and Madina were

used in this study. Samples for each variety were obtained

from at least three shops in Oman, and the varietal purity

was confirmed by a date variety expert at Sultan Qaboos

University. A conglomerate sample of 108 dates was taken

for each variety (n = 108 for each variety) and the sample

was imaged (single date images) with a color camera

(model: D3, Nikon, Japan, Resolution4,256 9 2,832

pixels). The date samples were illuminated with halogen

lamp (Visa tech, model SOLO 1600B) during imaging.

Then all images were converted into gray scale images

using Matlab software, and analyzed.

Feature selection

As the background of these images was deemed to have

little information regarding the classification of each date, a

thresholding operation was done in order to eliminate it

based on the maximization of the between-class variance.

The method is well known as Otsus method [7] and it is

briefly described next.

Starting by considering an image with background class

defined as wb, and object in class wo, with probabilities

P wb PT1

i1 f zi and P wo PL

iT fzi where

z denotes discrete image intensity, f(zi), i = 1, 2, , L is

the corresponding histogram normalized to have area equal

to one, and L is the number of distinct intensity levels; the

threshold T is found by maximizing r2

B

T

p wb lb ltot 2 p wo ltot

2where lb

PT1i1 zifzi,

lo PL

T zifzi and ltot PL

i1 zifzi.

Statistical features in the form of standardized moments

were to be considered [8]:

cn l

n

rn

E z m n

rn1

where Eis the expected value, r the standard deviation and

ln is the nth moment about the mean. For n = 3 one can

find skewness that measures the asymmetry of the proba-

bility density function of the random variable associated to

the image values. For n = 4 the value is known as kurtosiswhich also evaluates the shape of a distribution by quan-

tifying its peakness.

Because processing speed was important, out of the

mean, standard deviation, skewness, and kurtosis, we

decided to work with only three of them which would

eliminate approximately 25 % of the computations. Then,

for feature evaluation, a selection scheme with a backward

sequential search starting with a full feature set and

sequentially removing features was considered to be a

practical solution for the decision of what features to work

with. We decided to use a new feature selection scheme

that is based on fuzzy entropy measures with a similarityclassifier as presented by Luukka [9], which is briefly

described next.

The method starts by forming an ideal vector

vi = (v1(f1), , vi(ft)) that represents the class I as good as

possible by calculating the mean values of available vec-

tors in each class vi(fi) for i = 1, 2, , t features. After-

wards, the similarities Shx; vi between the sample x to beclassified and the ideal vectors v need to be calculated as:

Shx; vi 1

t

Xtr1

1 x fr vfrj j 2

In order to calculate the relevance of the features, fuzzy

entropy values are calculated with similarity values lA xj

as suggested by DeLuca and Termini [10]:

H1 A Xnj1

lA xj

log lA xj

1 lA xj

log 1 lA xj

3

where low entropy indicates high similarity values

and high entropy values are obtained otherwise.

G. Thomas et al.

123


3/6

Similarly, fuzzy entropy as suggested by Parkash et al. [11]

was used in this work which follows the expression:

H2 A Xnj1

sinplA xj

2

sinp 1 lA xj

2

1

4

Effect of motion

To validate the use of the classifier in motion conditions, a

simulation was conducted where the blurring degradation

caused by motion in one direction was implemented in the

frequency domain as [12]:

H u; v Te

puasin puaejpua 5

where u and v are normalized frequency samples, Te is the

camera exposure time, and a is the rate of movement in the

x axis direction: xo(t) = at/T. When t= T the image has

been displaced by a total distance a.

Results and discussion

Evaluation of features

Figure 1 illustrates three samples of each variety of dates.

The standardized moments were calculated only within the

pixels located within the date and not the background and

were computed for one date image at a time. From the

small image sample, it can be inferred that features such as

the mean value would work for the class Fard, since theylook darker than the other two classes. Figure 2 shows the

mean values m for 108 images of each class calculated with

the expression mean m = ltot. Note how in Fig. 2 the

values for class Fard do appear darker and dissimilar than

the other two.

We tested the four features (mean, standard deviation,

skewness, and kurtosis) and evaluated them according to

Eqs. (3) and (4). The selection process consistently elimi-

nated the values of kurtosis. With this result, the next

section elaborates on the design of the classifier using only

the mean, standard deviation, and skewness as inputs.

Classifier selection and classification results

As described in the previous section, three features were to

be used in the final classifier. Figure 3 shows a scatter plot

of the features and albeit three clusters are visible, there is

still some overlapping. A Lillilifors test [13] of the default

null hypothesis that the values of the different features

come from a Gaussian distribution passed the test con-

firming the Gaussianity of the clusters. Afterwards, a Bayes

classifier for Gaussian pattern classes under the condition

of a 01 loss function was implemented forming three

decision functions dj(x) for j = 1, 2, 3 of the form [12]:

dj x ln P wj

1=2 ln Cj

1=2 x mj T

C1j x mj h i

6

where mj = Ej{x} are the means of the feature grouped invector x, and Cj = Ej{(x2mj)(x2mj)

T}.

This classifier was tested using 80 samples for training

and 28 samples for testing. A 65.48 % of correct classifi-

cations were obtained this way, and in particular there were

only two misclassification for the class Madina yielding a

92.86 % of accuracy for this variety. When kurtosis was

used instead of skewness, a performance of 35.71 % was

obtained. This confirmed that the selection made by

Luukkas approach was correct.

Fig. 1 Gray scale images of date samples (first row Fard, second row

Khalas, third row Madina)

0 20 40 60 80 100

60

70

80

90

100

110

120Mean

Sample Index

Value

Khalas (blue)Fard (green)

Madina (red)

Fig. 2 Average gray level values of date samples

Dates varieties and effect of motion blurring

123


4/6

At this point, a two layer neural network was used with

hyperbolic tangent sigmoid transfer functions using Matlab

as the software platform for development. The training

algorithm used was the scaled conjugate gradient. In order

to define the number of hidden neurons N, neural networks

were trained 30 times with different numbers of hiddenneurons. The initial weights are calculated randomly by

Matlab and these averages would give an idea of the best

number of hidden neurons to be used. Because only three

features are used, the expectation was that the network

would not need many neurons. Table 1 shows the average

results after training the different networks 30 times with

respect to mean squared error (MSE) as well as time

required for training. Because of the small numbers of

neurons used, the training time was very short for all of

them.

From these results, it was decided that the number of

hidden neurons to use was five, not only because it had theminimum MSE value in Table 1 but also because of the

computational savings of not having to compute more

multiplications once the network had to perform classifi-

cations. Figure 4 shows the performance obtained using

such a network. The total number of epochs needed was 23,

for each class 65 out of 108 images were used for training,

best performance was obtained with MSE values of 3.81 %

for testing, 9.7 % for training and a needed training time of

0.78 s. Once trained, classifications were performed in less

than 0.025 s using an Asus Ee Slate tablet computer with

an Intel i5 1.33 GHz processor and 4 GB of RAM memory.

At this point the specified classification efficiency of

more than 90 % was achieved with a simple neural net-

work architecture to allow the option of a future imple-mentation in hardware rather than in a computer based

software solution. Thus, the solution was considered to be

fast enough to actually be implemented in a scenario where

a conveyor belt with continuous motion could be used.

Effect of motion

Figure 5 shows blurring simulated using Eq. (5) for a = 20

for the images shown in Fig. 1. Because a bright illumi-

nation set up was contemplated so that to eliminate any

possible shadows of the dates in bulk, faster shutting

camera times are expected and the blurring in Fig. 5 wasconsidered an extreme case.

In order to have an idea of the effects this blurring will

cause to the features, 5,000 images of size 256 9 256

formed by random numbers drawn from the distribution in

the Pearson system with specified l, r, c3, and c4, were

generated and blurred for different values of a. Figure 6

shows how these features changed as the blurring pro-

gresses. As the degradation function in (5) is a low pass

filter, blurring can be seen as a weighted averaging and no

changes in mean values are expected. The slight changes in

Fig. 6a correspond to the darkening of the beginning and

final values of the blurred images caused by discreteimplementation of the filter that viewed in the discrete time

domain, convolution values have this windowing effects.

The standard deviation is reduced as expected after low

pass filtering and skewness tend to go to zero, making the

samples more Gaussian as what we would see by averaging

random samples and explained by the central limit

theorem.

We used these features using Bayes classifiers in order

to identify the image blurring effects on classification.

60

80

100

120

15

20

25

30

35

-2

0

2

mean

standard deviation

skewness

KhalasMadinaFard

Fig. 3 Sctatter plot of the selected features for date samples

Table 1 Mean squared error (MSE) obtained while using a two layer

neural network with different number of hidden neurons

No. of neurons

n = 3 n = 5 n = 10 n = 15 n = 20

MSE (%) 12.73 11.04 12.27 13.62 13.79

Time (s) 1.0915 1.0130 1.0031 1.0441 1.0711

0 5 10 15 20 250

0.2

0.4

0.6

0.8

Epoch

MSE Train

Test

Fig. 4 Mean squared error (MSE) using a two layer neural network

with five hidden neurons. Matlab defines MSE as the measurement of

the networks performance according to the mean of squared

classification errors

G. Thomas et al.

123


5/6

Figure 7 shows the results when using both skewness and

kurtosis. As it can be seen, motion blurring at these rates

did not cause any major differences and it was further

confirmed that rejecting the kurtosis feature was correct as

indicated by the feature selection section. Similar two layer

networks as the one used in Fig. 4 were used for different

values of a, and Fig. 8 shows the consistent good results.

Results with accuracies similar to the ones found in

Pydipati [14] for detection of citrus disease were found

using a simpler neural network and reduced number of

features. Because of this simplicity, real time classification

under motion blurring caused by examining the dates on a

conveyor belt was deemed feasibly. Thus, we developed a

methodology to introduce motion via a degradation func-

tion implemented in the frequency domain to confirm the

robustness of the proposed method.

Acknowledgments We thank The Research Council (TRC) of

Sultanate of Oman for funding this study (Project No. RC/AGR/

SWAE/11/01-Development of Computer Vision Technology for

Quality Assessment of Dates in Oman).

References

1. A. Manickavasagan, G. Sathya, D.S. Jayas, N.D.G. White,

J. Cereal Sci. 47, 518 (2008)

Fig. 5 Motion blurred images of date samples (first row Fard, second

row Khalas, third row Madina)

0 5 10 15 20126

127

128

129

mean

(a)

0 5 10 15 200

2

4

6

stan

darddeviation

(b)

0 5 10 15 200

0.05

0.1

skeewnes

motion value a

(c)

Fig. 6 Results of motion blurring on selected features of date

samples

0 5 10 15 20

0.35

0.4

0.45

0.5

0.55

0.6

0.65

0.7

motion rate a

correctclassification

kurtosis

skewness

Fig. 7 Bayes classifier results for different motion rates

0 10 20 30 40 50 600

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

Epoch

MSE

a=4.8495

a=9.899

a=14.9495

a=20

Fig. 8 Mean squared error (MSE) obtained while using a two layerneural network with blurred images (results of test images)

Dates varieties and effect of motion blurring

123


6/6

2. S. Mahesh, D.S. Jayas, J. Paliwal, N.D.G. White, Sens. Instrum.

Food Qual. Saf. 5, 1 (2011)

3. M.I. Mladenov, S.M. Penchev, M.P. Dejanov, M.S. Mustafa,

Sens. Instrum. Food Qual. Saf. 5, 111 (2011)

4. S.R. Delwiche, M.S. Kim, Y. Dong, Sens. Instrum. Food Qual.

Saf. 5, 63 (2011)

5. FAO Statistics (2007), http://faostat.fao.org/site/342/default.aspx.

Accessed 24 May 2012

6. A.S. Al-Marshudi, Tropicultura 20, 203 (2002)

7. N. Otsu, IEEE Trans. Syst. Man Cybernet. 9, 62 (1979)

8. J.F. Kenney, E.S. Keeping, Mathematics of Statistics, Pt. 1, 3rd

edn (Princeton, NJ, 1962)

9. P. Luukka, Expert Syst. Appl. 38, 4600 (2011)

10. A. DeLuca, S. Termini, Inform. Cont. 20, 301 (1971)

11. O.M. Parkash, P.K. Sharma, Mahajan, Inf. Sci. 178, 2389 (2008)

12. R.C. Gonzalez, R.E. Woods, Digital Image Processing,, 2nd edn.

(Prentice Hall, NJ, 2001)

13. H.W. Lilliefors, J. Am. Stat. Assoc. 62, 399 (1967)

14. R. Pydipati, T.F. Burks, W.S. Lee, Trans. ASAE 48, 2007 (2005)

G. Thomas et al.

123
http://faostat.fao.org/site/342/default.aspxhttp://faostat.fao.org/site/342/default.aspx

classification of dates varieties and effect of motion blurring

Documents