Content uploaded by Lilibeth Coronel

Author content

All content in this area was uploaded by Lilibeth Coronel on Dec 06, 2018

Content may be subject to copyright.

708 The International Arab Journal of Information Technology, Vol. 15, No. 4, July 2018

Identification of an Efficient Filtering-

Segmentation Technique for Automated Counting

of Fish Fingerlings

Lilibeth Coronel1, Wilfredo Badoy2, and Consorcio Namoco3

1College of Science and Environment, Mindanao State University at Naawan, Philippines

2Department of Information Systems and Computer Science, Ateneo de Davao University, Philippines

3College of Industrial and Information Technology Mindanao, University of Science and Technology, Philippines

Abstract: The counting of fish fingerlings is an important process in determining the accurate consumption of feeds for a

certain density of fingerlings in a pond. Image processing is a modern approach to automate the counting process. It involves

six basic steps, namely, image acquisition, cropping, scaling, filtering, segmentation, and measurement and analysis. In this

study, two (2) filtering and two (2) segmentation algorithms are identified based on the following observations: the non-

uniform brightness and contrast of the image; random noise brought about by feeds, waste, and spots in the container; and the

likelihood of the image samples or application used by the different authors of the smoothing and clustering algorithms in their

respective experiments. Four (4) combinations of filtering-segmentation algorithms are implemented and tested. Results show

that combination of local normalization filter and iterative selection threshold yield a very high counting accuracy using the

measurement function such as Precision, Recall, and F-measure. A Graphical User Interface (GUI) is also presented to

visualize the image processing steps and its counting results.

Keywords: Digital image processing, filtering, segmentation, image normalization, threshold.

Received July 9, 2015; accepted February 3, 2016

1. Introduction

Fish fingerling has to be handled several times before

stocking into ponds or containers. Fingerling stocking

marks the beginning of a production cycle. It is among

the most delicate and stressful processes the fingerlings

go through in the course of production. The process of

stocking starts with the counting of fingerlings from

the hatchery, transporting them to the farm and, finally,

putting them into the pond. Traditional process of

counting fish fingerlings such as manual, volumetric,

and surface area methods are still adopted nowadays.

Problem like handling of fingerlings are very critical

since the fingerlings are weighted in volume in a

container and again count them manually. But today,

through the continuous advancement of information

technology, innovations and integration between

computer science and aquaculture facet is very

promising and remarkable. Image processing is a

rapidly growing area of computer science. It involves

six basic steps, namely, image acquisition, cropping,

scaling, filtering, segmentation, and measurement and

analysis.

Several studies have been conducted to automate the

counting of fish using image processing [2, 5, 6, 10].

Different filtering and segmentation algorithms have

been used for different purposes of the studies, namely,

counting accuracy, classification, behavioural aspects

of the fish, and the overall installation and setup of

image/video acquisition. Problems like error of

counting escalate as the number of fish increases; fish

sizes are unknown; and fish orientation differ.

Sometimes fish may not be segmented reliably,

lighting variations in acquiring image and changes in

water quality, thus causes an error in counting. In

addition, installations and the use of fragile equipment

can be problematic when used in remote locations.

In this study, we identify two filtering and two

segmentation algorithms based on the following

observations: the non-uniform brightness and contrast

of the image; random noise brought about by feeds,

waste, and spots in the container; and the likelihood of

the image samples or application used by the different

authors of the smoothing and clustering algorithms in

their respective experiments. Such filtering and

segmentation algorithms considered in this study

include Local Normalization filter [9], Median filter

[7], Iterative Selection threshold [8] and Minimum-

Error threshold [4].

The study also designed and developed a prototype

to automate the counting of fish fingerlings employing

the image processing steps with the identified

combinations of filtering and segmentation techniques

using java-programming language and an open source

image processing and analysis program.

Identification of an Efficient Filtering-Segmentation Technique for ... 709

2. The Research Method

Figure 1 shows the overall image processing system

model used in this study. Combinations of filtering and

segmentation algorithms are applied to identify the

efficient technique for counting the fish fingerlings.

2.1. Image Acquisition

In this step, images of tilapia fish fingerlings are

acquired using a Canon PowerShot A3200 IS digital

camera with 14.1 megapixels. The samples are placed

in a white plastic dishpan of size 48x12 cm with water

level of 4.45 cm in height. The size and age of the

samples are approximately 14-16 mm and 21-28 days

old.

Figure 1. Image processing system model used.

The position of the camera must be stable and the

distance from the dish is 60 cm in height with respect

to the dimension and areas inside the dishpan. The

images are taken between 1:00 to 3:00 in the afternoon

notwithstanding the lighting installation and setup.

Figure 2 shows the image acquisition setup used in the

implementation. The camera settings applied include:

ISO speed at ISO-80, F-stop of f/8, focal length of 5

mm with aperture of 2.968, and flash mode is set to

Off. The dimension of the image is 4320 x 3240 pixels.

The image file is in Joint Photographic Experts Group

(JPEG) format.

Figure 2. Image acquisition setup.

2.1.1. Input Image Representation

The image is denoted as two-dimensional function of

the form f(x, y). The amplitude of the image

f

at spatial

(plane) coordinates (x, y) is a positive scalar quantity

whose physical meaning is determined by the source of

the image. The image result has

W

rows and H

columns. The complete W x H digital image in a

compact matrix form is:

)1,1()1,1()0,1(

)1,1()1,1()0,1(

)1,0()1,0()0,0(

),(

HWfWfWf

Hfff

Hfff

yxf

The right side of this equation is by definition a digital

image. Each element of this matrix array is called

pixel.

2.2. Image Cropping

In this process, the image is cropped using ImageJ

built-in cropping function as seen in Figure 3.

Figure 3. Image cropping.

2.3. Image Scaling

After cropping, the image passes through the image

scaling using again the ImageJ built-in scaling function

as seen in Figure 4.

Figure 4. Image scaling.

2.4. Filtering

In the filtering step, two techniques are used separately

as filtering step in the image processing system.

2.4.1. Local Normalization

The local normalization [9] is a modern and efficient

filtering technique used for correcting non-uniform

(1)

710 The International Arab Journal of Information Technology, Vol. 15, No. 4, July 2018

illumination or brightness and eliminates the effects of

uneven noise in an image. The local normalization of

f(x,y)

is computed as:

( , ) ( , )

( , )

( , ) f

f

f x y m x y

xy

g x y

Where f(x,y) is the input image, mf (x,y) and σf (x,y)

represents the estimation of the local mean and

variance of f(x,y) and g(x,y) is the output filtered

image. The local mean and variance of the image are

estimated by a recursive Gaussian filter. The

parameters of the algorithm are the sizes of the

smoothing window σ1 and σ2 which value is larger than

σ1 that controls the estimation of the local mean and

variance.

2.4.2. Median Filter

The median algorithm [7] is the simplest and widely

used median filtering that normally reduces random

noise in an image. The values of the pixel in the

window are stored and the median – the middle value

in the sorted list (or average of the middle two if the

list has an even number of elements)-is the one plotted

into the output image. The median filtered image g(x,y)

can be obtained from the median pixel values in a

neighborhood of (x,y) in the input image f(x,y) , as

defined by the following formula:

1

1

1

1

),(),( i j

jyixfmedianyxg

2.5. Segmentation

Similarly, two techniques are used separately as

segmentation step in the image processing system.

2.5.1. Iterative Selection Threshold

The simplest of all the thresholding techniques is the

iterative selection method [8]. The method models the

gray level distribution in an image as mixture of two

Gaussian distributions representing, the background

and foreground region. The threshold is computed as:

Tn=mf,0 +mb,0

2

Where at iteration n, a new threshold Tn is computed

using the average of the foreground mf,0 and

background mb,0class means. On each iteration, the

mean gray level for all pixels greater than

T

is

determined, and is denoted asG1. The mean gray level

for all pixels lesser than or equal to T is also

determined, and is denoted asG2. Iteration terminates

when the changes |Tn¯Tn+1 becomes sufficiently small.

2.5.2. Minimum Error Threshold

The minimum error threshold algorithm [4] is based on

the assumption of object and background pixels gray

level values in the image being normally distributed.

Normal distributions are defined by their means µi,

standard deviations σi, and a priori probabilities Pi. The

background and foreground represents two different

classes (i=1, 2) and a given threshold T. The minimum

error threshold can be computed by minimizing the

criterion function J(T) calculated as:

1 1 2 2

1 1 1 2 2 2

log log

2 log log

( ) 1 2 P T T P T T

P T P T P T P T

JT

This is applied since some of the images have non-

uniform brightness or poor intensity condition with

fishes as object of interest.

2.6. Measurement

Measurement of the object of interest or the fingerlings

thru defining the size, circularity ranges of the

fingerlings as pixels and its height-to-width-ratio is

implemented. The output of this measurement method

is the summary of the total number of fish fingerlings

identified. The parameters of this method are the sizes

and circularity that specifies the range pixel values

inside the object and its shape. The parameters may

vary depending on the age of the fingerlings.

The number of pixels it contains defines the size of

an object. A patch P consisting represents each object

of a list of lines l, the number of pixels n is given by:

PLl se lxlxn

The circularity Cr specifies the object-based shape

measurement calculated by the formula:

2

4

rA

C

Where A is the area and

is the perimeter.

In determining the object of interest, the Height-To-

Width Ration (HTWR) is applied represented by a

patch P is given by:

max min

max min

y i P y i P

x i P x i P

HTWR

Which helps identify objects that are either too long

compared to their height or too tall compare to their

breadths. Thus, if either of the following conditions for

patch P is satisfied,

HTWR(P)<HTWRmin

HTWR(P)>HTWRmax

P is classified as noise and taken off from the

PatchList.

2.6.1. Identifying Combinations of Filtering and

Segmentation Techniques

In identifying Combinations C for the experiments,

where F stands for filtering (Local Normalization,

Median filter) and S stand for segmentation (Iterative

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Identification of an Efficient Filtering-Segmentation Technique for ... 711

Selection threshold, Minimum Error threshold), the

product F x S is the set of all pairs (f,s) where f denotes

filtering techniques such as Local Normalization and

Median, f ϵ F and s denotes segmentation techniques

such as Iterative Selection and Minimum Error, s ϵ S.

The groupings must satisfy the following rules for each

of the image processing model:

1. A group must have only one filtering and one

segmentation technique.

2. The filtering technique must come first before

segmentation technique.

These combinations are evaluated based on the

accuracy of counting the fish fingerlings, namely,

Combination A (Local Normalization and Iterative

Selection), B (Local Normalization and Minimum

Error), C (Median and Iterative Selection) and D

(Median and Minimum Error), respectively.

2.6.2. Counting Evaluation

Four combinations of filtering and segmentation

algorithms are compared and evaluated thru calculating

the following information retrieval measures, namely,

Precision, Recall, and F measure. The evaluation is

widely used in other studies in terms of image analysis

[1, 2, 3]. The Precision (P), Recall (R), and F-measure

are calculated by:

fptptp

P

fntptp

R

F measure =2* P*R

P+R

where True Positives (TP) represent the number of fish

fingerlings correctly identified as fish fingerlings,

False Positive (FP) represents incorrectly identified by

the method as a fish fingerlings such as noise in the

image and False Negative (FN) represents fish

fingerlings that are not identified as fish fingerlings but

are existed.

Furthermore, the best F measure among the four

combinations is compared with the actual number of

fish fingerlings. That is, the level of closeness of

measurements of the total number of fish fingerlings to

that of the actual (true) number of fish fingerlings. The

computed F measure are between 0 and 1. A higher

value of F measure indicates a higher classification or

clustering quality and lower error rates or

misclassification of fish fingerlings.

3. Experimental Results

The image processing system is implemented as a

plugin to an image processing software and analysis

tool (ImageJ, [7]) employing the four identified

combinations of filtering and segmentation techniques

to automate the counting of fish fingerlings.

Figures 5, 6, and 7 show the visualization results of

the image processing system employing the

Combination A technique. Figure 5 shows the sample

acquired JPEG format image of Tilapia fish fingerlings

with the dimension of 4320x3240 pixels used as input

image. The input image is cropped to approximate the

region of interest. The dimension of the image is

reduced to 3180 x 3180 pixels as shown in Figure 6.

The cropped image is rescaled to half of its

dimension to downsize and classify the pixel values

that surrounds the image as shown in Figure 7. The

dimension of the image is reduced to 1590x1590

pixels. The scaled image is filtered according to the

parameters set. The filtered image is segmented to

cluster the object of interest from its background

region. The segmented image is then processed to

measure the object of interest being identified.

Figure 8 shows the visualization results of the image

processing systems employing Combination A. The

image results show significantly very high in noise

reduction and feature identification thus generating

remarkable counting results. Combination B shows

significantly very high in noise reduction but very poor

in feature identification thus generating very poor

counting results. This is shown in Figure 9. Figures 10

and 11 shows significantly poor in noise reduction and

relatively poor in feature identification thus also

generating poor counting results for both Combinations

C and D. Figure 12 shows the ImageJ Graphical User

Interface (GUI) of the automated system using

Combination A.

The parameters defined are based on the actual size

of the fish fingerling samples, the non-uniform

brightness and contrast of the image; random noise

brought about by feeds, waste, and spots in the

container. Combinations A and C requires two input

parameters since these combinations used Local

Normalization as the filtering method. Such parameters

includeσ1=2, σ2=50, size=80-300pixels, and

circularity=0.09-1.0. While Combinations B and D

only requires radius=2 as input parameter with the

same size and circularity values.

Table 1 shows the average measurement result of

the image processing system employing the four

different combinations of the filtering and

segmentation techniques. The experiment considered 2

groups of images, each group having 5 images. Group

A contains 50 Tilapia fish fingerlings in each image

and Group B has 100 fish fingerlings in each image.

The result shows that Combination A has the

highest Precision, Recall, and F measure values as

compared to other combinations of the filtering and

segmentation techniques.

(10)

(11)

(12)

712 The International Arab Journal of Information Technology, Vol. 15, No. 4, July 2018

Figure 5. Sample tilapia fish fingerling.

Figure 6. Cropped image.

Figure 7. Scaled image.

a) Local Normalization filter. b) Iterative Selection threshold.

Figure 8. Image result using combination A.

Table 1. Average measurement results of the four combinations of

Filtering and Segmentation techniques.

Images

Techniques

Measures (%)

Precision

Recall

F measure

Group A

Combination A

99.59

98.41

98.99

Combination B

NaN

23.60

NaN

Combination C

89.20

49.20

63.15

Combination D

85.92

33.65

48.06

Group B

Combination A

100

97.40

98.68

Combination B

NaN

11.60

NaN

Combination C

91.26

48.40

62.90

Combination D

84.46

34.60

48.83

a) Local Normalization filter. b) Minimum Error threshold.

Figure 9. Image result using combination B.

a) Median filter b) Iterative Selection threshold.

Figure 10. Image result using 1Combination C.

a) Median filter. b) Minimum Error threshold.

Figure 11. Image result using combination D.

Figure 12. ImageJ GUI of the automated system.

Table 2. Detailed experimental results comparing the combination

A and the manual counting process.

Image no.

Automated Counting System using

Combination A technique

Manual Counting

process

No. of

Tilapia

Fingerlings

Measures

No. of

Tilapia

Fingerlings

Measures

TP

FP

FN

P

R

F measure

TP

FP

FN

P

R

F

measure

Group A

( with 50

Fingerlings)

1

50

0

0

1

1

1

50

0

0

1

1

1

2

49

0

1

1

0.98

0.98989899

50

0

0

1

1

1

3

49

0

1

1

0.98

0.98989899

50

0

0

1

1

1

4

50

0

0

1

1

1

50

0

0

1

1

1

5

48

1

2

0.979591837

0.96

0.96969697

50

0

0

1

1

1

Average

0.995918367

0.984

0.98989899

Average

1

1

1

Group B

(with 100

Fingerlings)

6

97

0

3

1

0.97

0.984771574

100

0

0

1

1

1

7

95

0

5

1

0.95

0.974358974

100

0

0

1

1

1

8

97

0

3

1

0.97

0.984771574

100

0

0

1

1

1

9

100

0

0

1

1

1

100

0

0

1

1

1

10

98

0

2

1

0.98

0.98989899

100

0

0

1

1

1

Average

1

0.974

0.986760222

Average

1

1

1

Identification of an Efficient Filtering-Segmentation Technique for ... 713

This means that Combination A has the highest

percentage of correctly identified fish fingerlings, the

lowest percentage of incorrectly identified fish

fingerlings and the lowest percentage of noise

identified. It is observed that the Combination B yields

insignificant results in which Precision and F measure

result are Not-a-Number (NaN). It can be seen from

Table 1 that Combinations C and D cannot outperform

the accuracy of Combination A.

On the other hand, the efficiency of the local

normalization technique to correct non-uniform

lighting and the reduction of noise in an image are very

significant. The Iterative Selection technique also

achieves significant results in feature identification.

It can also be observed from Table 2 that the

variance of counting results between Combination A

and the manual counting are 0.20%, 2.10% and 1.17%

in average Precision, Recall and F measure values,

which are very minimal. These measurements indicate

that result of the automated system is very close to the

manual counting results.

The automated counting system results show that

the total number of correctly identified fish fingerlings

(tp) are very high in the two groups of images but it is

observed that as the number of fish fingerlings

increases, the number of fish fingerlings that are

present but are not actually counted (as if they did not

existed) (fn) also increases. Moreover, the automated

counting system using Combination A obtained the

value of zero (0) in terms of the number of incorrectly

identified as fish fingerlings (fp) such as noise, hence it

is very close to that of the manual counting.

Furthermore, the rate of time in counting the fish

between automated and manual system is also

compared as seen in Table 3. The automated system is

tested on an Intel Core i5 processor with 4Gb of

memory. The time measurement indicates that

automated system significantly performs best

compared to manual counting process.

4. Conclusions

Combination A (Local Normalization and Iterative

Selection) provides significantly very high in

correcting non-uniform lighting in an image, noise

reduction and feature identification compare with other

combinations of filtering and segmentation techniques.

In terms of counting accuracy, Combination A

obtained an average Precision, Recall and F measure

of as high as 99.80%, 97.90% and 98.83% which

outperformed other combinations, respectively.

Moreover, with the automated system, manual-

counting delays can be resolved. Mortality rate of the

fish fingerlings also decreases, thus, providing an

increase in production of fish fingerlings among the

growers and aquatic biological experts.

Table 3. Time measurement results between automated counting

system using combination A and manual counting process.

Image no.

Automated Counting System

using Combination A technique

Manual Counting

process

Count

Time (sec)

Count

Time (sec)

Group A

(with 50

Fingerlings)

1

50

0.526

50

28

2

49

0.503

50

26

3

49

0.488

50

25

4

50

0.478

50

24

5

49

0.494

50

23

Group B

(with 100

Fingerlings)

6

97

0.501

100

55

7

95

0.446

100

53

8

97

0.513

100

51

9

100

0.575

100

50

10

98

0.512

100

49

5. Recommendation

For future work, it is suggested to further enhance the

identified technique to improve the counting efficiency

of the automated system with higher number of

samples that would not just provide counting statistics

but as well as identify the type of fingerlings.

Implementation of connected and independent

components algorithm is also recommended for

accidentally connected or overlapping fingerlings. The

algorithms used in the study may also be enhanced for

possible applications in other non-aquatic and non-

biological samples.

References

[1] Alhutaish R. and Omar N., “Arabic Text

Classification using K-Nearest Neightbour

Algorithm,” The Internation Arab Journal of

Information Technology, vol. 12, no. 2, pp. 190-

195, 2015.

[2] Cadieux S., Michaud F., and Lalonde F.,

“Intelligent System for Automated Fish Sorting

and Counting,” in Proceedings of International

Conference on Intelligent Robots and Systems,

Takamatsu, pp. 1279-1284, 2000.

[3] Faustino G., Gattass M., Rehen S., and de Lucena

C., “Automatic Embryonic Stem Cells Detection

and Counting Method in Fluorescence

Microscopy Images,” in Proceedings of IEEE

International Symposium on Biomedical

Imaging: from Nano to Macro, Boston, pp. 799-

802, 2009.

[4] Kittler J. and Illingworth J., “Minimum Error

Thresholding,” Pattern Recognition, vol. 19, no.

1, pp. 41-47, 1986.

[5] Labuguen R., Volante E., Causo A., Bayot R.,

Peren G., Macaraig R., Libatique N., and

Tangonan G., “Automated Fish Fry Counting and

Schooling Behavior Analysis Using Computer

Vision,” in Proceedings of IEEE 8th International

Colloquium on Signal Processing and its

Applications, Melaka, pp. 255-260, 2012.

[6] Morais E., Campos M., Padua F., and Carceroni

R., “Particle Filter-Based Predictive Tracking for

Robust Fish Counting,” in Proceedings of 18th

714 The International Arab Journal of Information Technology, Vol. 15, No. 4, July 2018

Brazilian Symposium on Computer Graphics and

Image Processing, Brazil, pp. 367-374, 2005.

[7] Rasband W., ImageJ, U. S. National Institutes of

Health, Bethesda, Maryland, USA,

http://imagej.nih.gov/ij/, Last Visited, 2014.

[8] Ridler T. and Calvard S., “Picture Thresholding

Using an Iterative Selection Method,” IEEE

Transactions on Systems, Man, and Cybernetics,

vol. 8, no. 8, pp. 630-632, 1978.

[9] Sage D., “Local Normalization.” Biomedical

Image Group. April 2009. Online. Nov. 6, 2011.

http://bigwww.epfl.ch/sage/soft/localnormalization/,

Last Visited, 2014.

[10] Toh Y., Ng T., and Liew B., “Automated Fish

Counting Using Image Processing,” International

Conference on Computational Intelligence and

Software Engineering, Wuhan, pp. 1-5, 2009.

Lilibeth Coronel completed her

Master’s Degree in Information

Technology at Mindanao University

of Science and Technology,

Philippines last 2014 and Bachelor’s

Degree in Computer Science at

AMA Computer College,

Philippines last 2001. Presently, she is an assistant

professor in the Department of Information

Technology at Mindanao State University-Naawan

Campus, Philippines and at the same time ICT Unit

Head of the Campus.

Wilfredo Badoy received an

Electronics Engineering degree at

Mindanao Polytechnic State College

in 1995, his MS Information

Technology at Ateneo de Davao

University in 2009. He is currently

finishing his Ph.D. in Computer

Science at Ateneo De Manila University. He has

worked with various schools in Northern and Southern

Mindanao for more than 15 years. His interests are in

Artificial Intelligence, Affective Computing, and

Computer Simulation. He has published researches in

journals and conference proceedings.

Consorcio Namoco completed his

Doctor of Engineering from Kyoto

Institute of Technology, Kyoto City,

Japan last 2012. His research

interests are in the fields of metal

forming, computer simulation,

materials processing, industrial

technology and information technology education.

Presently, he is a full professor and the vice chancellor

for academic and student affairs, University of Science

and Technology in Southern Philippines, Cagayan de

Oro City, Philippines. He also serves as editorial board

member and peer reviewer to various local and

international research journals.