Automated Vision System for Cutting Fixed-weight or Fixed-length

Frozen Fish Portions

Dibet Garcia Gonzalez

, Nelson Alves

, Ricardo Figueiredo

, Pedro Maia

and Miguel Angel Guevara Lopez

1,4

Computer Center Graphics, Minho University, Guimarães, Portugal

Gel Peixe, Loures, Portugal

Sociedade Industrial de Robótica e Controlo, S. A., Aveiro, Portugal

Centro Algoritmi, Minho University, Guimarães, Portugal

Keywords:

Computer Vision, 3D Reconstruction, Automation, Frozen Fish Automatic Cut, Optimization, Automated

Optical Inspection.

Abstract:

The increase in ﬁsh demand poses a challenge to the food industry that needs upgrades both for: (1) offering

new and diverse products and (2) optimizing it’s processing line throughput and at the same time guaranteeing

the ﬁshes’ quality and appearance. This work presents an innovative computer vision system, to be integrated

into an automatic frozen ﬁsh cutting production line. The proposed system is able to perform the 3D recon-

struction in real time of every frozen ﬁsh, and allows to: (1) identify and automatically separate head bone,

body and tail parts of the ﬁshes; and (2) estimate with high accuracy where cut the ﬁshes’ body to produce

the wanted slices, according to requirements (parameters) of weight or width previously deﬁned. The exper-

imental and statistical results are very promising and show the viability of the developed system. As main

contribution (novelty) this new method is able to estimate automatically and with high precision the weight of

the part corresponding to the body of the ﬁshes and thus optimizing the cut of the ﬁsh slices. With this, we

expect to achieve a signiﬁcant reduction of ﬁsh losses.

1 INTRODUCTION

Fishing is one of the sources of food, nutrition, ﬁ-

nancial income and livelihoods for hundreds of mil-

lions of people around the world. The supply of ﬁsh

at worldwide level reached a record high of 20kg

per capita in 2014 (Food and of the United Nations,

2016). This increase in ﬁsh demand poses a chal-

lenge to the industry whose current approaches for

ﬁsh cutting are mainly based on manual work and cuts

are made using only ﬁxed dimensions. On the other

hand, consumers want to buy ﬁsh by weight, without

scales and with a great visual appearance (freshness).

This demand, imposes a challenge to the food indus-

try that needs upgrades both for: (1) offering new

and diverse products and (2) optimizing it’s process-

ing line throughput and at the same time guaranteeing

the ﬁshes’ quality and appearance.

Currently, food processing, as it is the case of ﬁsh,

tends to be carried out in industrial and automated en-

vironments (Booman et al., 2010). Currently, automa-

tion is more and more a reality due to the well known

advantages. The Automatic (Automated) Optical In-

spection (AOI) algorithms and methods are currently

part of the "automation" bundle and allows for preci-

sion, repeatability, and uninterrupted periods of work

during all the year.

This work presents a new computer vision system,

to be integrated into an automated processing line,

able to the three dimensional (3D) reconstruction of

frozen ﬁsh (in real time), and automatically segment

the ﬁshes into three (deﬁned) parts (i.e. head bone,

body and tail). Also, the developed system will esti-

mate with high accuracy where (attending to an estab-

lished reference coordinate system) to cut the ﬁshes

slices according to requirements (parameters) such as

weight or width.

The AOI-based method proposed here is a tailor-

made solution developed for a speciﬁc client, a rep-

resentative player of the Portuguese and international

ﬁsh industry. This new system will allow to optimize

the cutting process, replacing the today current sce-

nario where a human operator is responsible for cut-

ting the frozen ﬁsh in a pre-determined and ﬁxed di-

Gonzalez, D., Alves, N., Figueiredo, R., Maia, P. and Lopez, M.

Automated Vision System for Cutting Fixed-weight or Fixed-length Frozen Fish Portions.

DOI: 10.5220/0007482407070714

In Proceedings of the 8th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2019), pages 707-714

ISBN: 978-989-758-351-3

707

mension while at the same time inspects it for "de-

fects" (e.g. if it is the "head bone" present and visi-

ble?).

2 PREVIOUS WORK

The computer vision technologies applied to the in-

dustry cover a great variety of solutions for image

acquisition like: visible/near-infrared, computed to-

mography, X-ray (Kelkar et al., 2015), magnetic res-

onance imaging, laser scanner (Kelkar et al., 2011) in

2D and 3D space (Mathiassen et al., 2011). The es-

timation of weight using computer vision techniques

implies the determination of density. Magnetic res-

onance imaging, computed tomography, laser scan-

ners,and even X-ray are imaging modalities that have

been frequently employed for this kind of task.

An overview of some of the AOI-based developed

solutions for "food inspection", in speciﬁc, focused

to compute (estimate) parameters, such as the weight

and density are presented to follow.

In (Viazzi et al., 2015), computer vision tech-

niques are used to estimate the weight of a complete

ﬁsh. The study uses regression techniques to ﬁnd the

best model that estimates weight using features ex-

tracted from images. The computer vision algorithm

includes four steps. First, detect the region that en-

closes the ﬁsh. Second, the ﬁsh is segmented using

the "Otsu adaptive thresholding" algorithm. Next, it

is removed the tail ﬁn using a shape analysis method.

Finally, some features like height, length, and area are

extracted with and without including the tail ﬁn. As

conclusions, authors state both for: 1) the area pa-

rameter is sufﬁcient and more robust to estimate the

weight of a ﬁsh; and 2) the tail ﬁn negatively inﬂu-

ences the weight estimation. Another approach tries

to establish a relation between the projected area of

sushi shrimp, captured using an RGB camera, and its

corresponding weight (Poonnoy and Chum-in, 2012).

In (Mortensen et al., 2016) for the use case of

broiler chickens weight estimation, it is described a

system composed of a weighing platform (used to

give a reference weight) and a RGB-D sensor (Mi-

crosoft kinect), where authors developed an algorithm

including the following steps: (1) segment the broil-

ers using watershed and region growing techniques

in a previously ﬁltered gray-scale image (combining

a Gaussian ﬁlter and morphological operators); (2)

the extraction of features like projected area, width,

perimeter, radius eccentricity, volume and others from

each segmented broiler; and (3) weight prediction of

each broiler based on the extracted features using ar-

tiﬁcial neural networks. A convex hull and numeri-

cal integration are used to obtain an approximated 3D

model of the broilers.

In (Adamczak et al., 2018) it is used a white light

(3D scanning) technology to determine the weight of

the chicken breasts.

Eggs volume estimation based on computer vi-

sion techniques was explored by (Soltani et al., 2015)

with the objective to extract the eggs major and minor

diameters. Two methods for estimating the volume

were tested: 1) a Mathematical model based on Pap-

pus theorem; and 2) Artiﬁcial Neural Networks. As

result, it was observed that the developed mathemati-

cal model achieved better results..

3 AUTOMATIC VISION SYSTEM

In order to support the development of the proposed

method, we built an experimental prototype (a stand

alone machine) comprising: 1) a conveyor belt that

moves at a speed of 0.5 m/s, which can transport

ﬁshes of different sizes and volumes; 2) a precision

scale; 3) two "Entry-Level 3D Laser Line Proﬁle Sen-

sors" - Gocators 2150 (Technologies, 2018) in an op-

posite layout placed at a distance of one meter (Fig-

ure 1). The Gocator 2150 is a 3D smart sensor / cam-

era, an all-in-one solution that lets factories to im-

prove efﬁciency in product validation; and 4) an in-

dustrial PC.

With this, each ﬁsh is weighted using the scale

and its total weight (T

) is recorded. Then, the ﬁsh

is scanned using the Gocators sensors, resulting two

line proﬁles (corresponding to both sides of the ﬁsh;

top and bottom). The computed proﬁles are then inte-

grated generating a precise 3D ﬁsh model using a ﬁne

tuning developed computer vision algorithm.

Figure 1: Hardware prototype of the acquisition module,

including sensors that emits and capture lines of laser that

are projected in the ﬁsh body over the conveyor belt.

Taking as input the generated 3D model, we estimate

the individual weight of a voxel (V

) dividing the total

weigh of the ﬁsh (T

) by the total number of the com-

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

708

puted voxels. A voxel is considered the minimum unit

of a 3D object (the ﬁsh in this case) with dimensions

w, l and h respectively. Additionally, it is computed

the total volume (T

) as the sum of all the volumes of

the voxels. The ﬁsh density (ρ) is calculated dividing

the total volume (T

) by the total weight(T

The cameras reference system is used as the global

coordinates system. Therefore, the two produced pro-

ﬁles are identiﬁed as the top and the bottom respec-

tively. The x axis coincides with the width (w), the z

axis with the height (h) and the y axis with the length

(l) of the ﬁsh.

The built 3D model is then segmented in three

parts: the head bone, body and tail ﬁn models. The

head bone is considered waste and body and tail ﬁn

are separated to be processed before introducing these

into the cutting system machine. In the next step, the

3D model of the body part is used as input to deter-

mine the optimal coordinates of slices for each in-

dividual ﬁsh, taking into account the previously se-

lected parameters (length or weight). The description

of the mechanical part of the cutting system machine

will not be discussed here, due to, it is not an objective

of this work.

3.1 Data Acquisition and

Pre-processing

For the development of this work, we used an exam-

ple dataset comprising 200 ﬁshes of Merluccius mer-

luccius with different sizes, supply by the "Gel Peixe"

company (http://gelpeixe.pt). The used sensors / cam-

eras (Gocator 2150) capture information such as: in-

tensity levels and depth proﬁles. However, we used

only, as input data, the top and bottom depth proﬁles

captured by the sensors, which are acquired when the

ﬁshes are moving through the conveyor belt, inside

of the sensors ﬁeld of view (the y axis). Cameras

are previously calibrated and aligned, guaranteeing

that the extrinsic parameters between both cameras

are known. An encoder is used to trigger the cameras

in a synchronized way. With this, we have a resolu-

tion that allows to capture voxels with the following

dimensions in mm: 1x1x0.6 (x,y,z) respectively.

The pre-processing step aims to remove every in-

formation that is not belong to the ﬁsh, such as the

foreign objects and the conveyor belt among others.

The foreign’ objects subtraction consists in the ap-

plication of a thresholding process to both proﬁles

(top and bottom) in which are correlated the width of

the conveyor belt and the maximum width of the ﬁsh.

The threshold was selected heuristically taking into

account the ﬁsh dimensions. It allows the selection

of a correct ROI (Region of Interest), which is criti-

cal for the execution of the next steps. The Figure 2

shows the measured proﬁle from each camera and the

manual threshold selection.

Figure 2: Example of proﬁles: Top (red) and Bottom (blue).

Heuristically selected thresholds (enclosed in green circles).

While the ﬁsh is crossing the ﬁeld of view of the cam-

eras, the proﬁles are captured and stored in order to

form two gray-scale images (Figure 3). These images

still need a ﬁltering step in order to remove the con-

veyor belt and other noise related to the proﬁle acqui-

sition.

(a) 2260gr (b) 2260gr

Figure 3: Bottom (a, c) and top (b, d) images without the

ﬁltering step

For this task, we employ mathematical morphology

operators combined with logical operations, as pre-

Automated Vision System for Cutting Fixed-weight or Fixed-length Frozen Fish Portions

709

sented to follow. These steps are applied to the bot-

tom ( f

) and top ( f

) two dimensional (2D) gray-scale

images.

1. Adjust the coordinate system.

(x,y) = f

(x,y) + min( f

(x,y)) (1)

(x,y) = f

(x,y) + min( f

(x,y)) (2)

where min( f

(x,y)) calculates the minimum (co-

ordinates in the z axis) value of the bottom proﬁle.

2. Image binarization

(x,y) =



1 if f

(x,y) 6= 0

0 if f

(x,y) = 0

(3)

(x,y) =



1 if f

(x,y) 6= 0

0 if f

(x,y) = 0

(4)

3. Binary mask

m(x,y) = b

(x,y) ∨ b

(x,y) (5)

4. Improving the binary mask. It is applied a mor-

phological opening operation with structuring el-

ement K to m(x,y), with this are removed small

(noise) particles.

m(x,y) = (m(x,y)  K(a,b)) ⊕ K(a, b) (6)

5. Gray image reconstruction

(x,y) = f

(x,y) × m(x, y) (7)

(x,y) = f

(x,y) × m(x, y) (8)

6. Blob (ﬁsh) detection and image cropping. In this

step we ﬁnd the regions of interest (the ﬁsh) in

and f

. Then, the bounding boxes (rectangles)

that enclose each blob are superimposed in order

to form a new rectangle that is used in the crop-

ping process of f

and f

. An example can be

seen in Figure 4.

7. A Gaussian ﬁltering is applied to reduce noise,

which allows to smooth the previously recon-

structed gray image.

The Figure 5 shows the image evolution since step

1 to step 7.

Figure 4: Rectangle estimation used in the cropping process

of f

and f

Figure 5: Image processing work ﬂow (steps).

3.2 3D Model Reconstruction and

Density Estimation

The ﬁltered top and bottom images are used as in-

put to build the 3D model. The corresponding row in

both images are used in the building of L new images

containing the edges (contours) of the region (area)

of the ﬁsh in each particular image (plane) (see Fig-

ure 6). Unconnected points are linked using a cubic

spline algorithm to close the ﬁsh area in each plane.

The number of pixels enclosed by this area is N

i=1..L

The total area (in mm

) is calculated using the equa-

tion:

i=1..L

= N

∗ w ∗ h (9)

The volume in mm

between two continuous im-

ages planes is computed as:

i=1..L

= A

∗ l (10)

The total ﬁsh volume (T

) in mm

is:

∑

i=1..L

(11)

Knowing the weight (T

) of each ﬁsh and its vol-

ume (T

) we can estimate its density ρ as:

ρ =

(12)

Finally, is calculated the weight associated to the

image as:

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

710

i=1..L

= V

∗ ρ (13)

Figure 6: Proﬁle samples.

3.3 Fish Segmentation

The cutting process of frozen ﬁsh always involves a

ﬁrst stage of bone cutting (i.e. the head bone). The

detection of the exact position where to cut the head

bone (L

bone

) is a trial and error process when done

manually. An experienced operator can make 2 to 3

cuts on the same ﬁsh before to achieve a ﬁrst slice of

the ﬁsh with the expected (optimal) quality. To extract

the correct place (position) for cutting the head bone,

we developed a simple algorithm that takes as input

the images obtained in the step 7, as it can be observed

in the sub-section 3.1 (see Figure 5).

To automatically identiﬁes the coordinates for cut-

ting the head bone, we developed an algorithm that

consists in moving a segment of line using an inclina-

tion angle of 30

◦

from the upper left edge downwards

in both images (top and bottom). This procedure is

repeated until (at least) the segment of line intersect

one of the pixels belonging to the ﬁsh, in any one of

the two images (top or bottom). The length of the tail

ﬁn (L

tail

) depends on the size of the ﬁsh and the opti-

mization strategies (i.e. the length that was previously

deﬁned by the costumer). With this, the body (L

) is

considered the part (portion) between head bone and

tail ﬁn (space between L

bone

and L

tail

, which is the

"useful space" for cutting the slices of the ﬁsh. The

Figure 7 shows two segment of lines, which identiﬁed

(divide) the head bone, body and the tail ﬁn section of

the ﬁsh.

3.4 Fish Cut Optimization Strategies

Once segmented the body of the ﬁsh, the next step is

related to computing the coordinates where its will

be produced the cuts for obtaining the optimized

ﬁsh slices. As required by the costumer, the ﬁsh

can be processed following one of two well-deﬁned

strategies: 1) Fixed-Length, using as input a selected

length, between a range of values to cut the ﬁsh slices;

and 2) Fixed-Weight, using as input a selected weight

to cut the ﬁsh slices.

The Fixed-Length strategy allows a length range

between min

length

and max

length

to apply the cuts.The

(a) 2260gr (b) 1880gr

Figure 7: Bone estimation and tail ﬁn section. The tail

length is ﬁxed to 3cm.

goal is to minimize the waste, therefore, it can be in-

tended as the minimization of the distance between

the last slice coordinates and the start-off position of

the tail ﬁn (L

tail

). For this purpose, we use the algo-

rithm 1. In our setup, we consider the size of the ﬁsh

as having L mm, due to the ﬁsh’s proﬁles coordinates

are acquired with a distance of 1 mm in the sense of

the y axis.

The algorithm 2 describes the Fixed-Weight strat-

egy.

Automated Vision System for Cutting Fixed-weight or Fixed-length Frozen Fish Portions

711

Algorithm 1: Fixed-Length strategy algorithm. The IF

condition allows adding 1mm more in another to consider

the loss produced by the saw.

Result: A vector with all the cuts positions

(positions)

= L − L

bone

− L

tail

;

cuts_number = 0;

rest_space = L

− cuts_number ∗ min_length;

while (rest_space > max_length) do

cuts_number + +;

rest_space =

− cuts_number ∗ min_length;

end

adding = rest_space/cuts_number;

real_cut_width = min_length + adding;

start = L

bone

;

positions = [];

for (i = 0;i < cuts_number;i + +) do

position_to_add = 0;

if (real_cut_width + 1 > max_length then

position_to_add = max_length;

else

position_to_add = real_cut_width + 1;

end

Insert position_to_add into the vector

positions;

start+ = position_to_add;

end

Algorithm 2: Fixed-Weight strategy algorithm.

Result: A vector with all the cuts positions

(positions)

start = L

bone

;

acc_weight = 0.0;

positions = [];

for (i = L

bone

;i < L

bone

+ L

;i + +) do

acc_weight+ = V

∗ ρ;

if (acc_weight >= desired_weight) then

position_to_add = i − start;

Insert position_to_add into the vector

positions;

start = i;

acc_weight = 0.0;

end

4 EXPERIMENTAL SETUP /

VALIDATION

In order to validate and verify the developed meth-

ods, we use the Merluccius merluccius dataset, a set

of ﬁshes (with different weights), which were contin-

uously scanned, and volume of every ﬁsh in each it-

eration was computed. Then, statistical descriptors,

such as, the mean and standard deviation were ex-

tracted. An example, using one ﬁsh can be observed

at Figure 8, which shows the volume proﬁles of a ﬁsh

(weighting 1880gr) computed 200 times in different

positions, using the Equation 10. The Table 1 shows

the statistical descriptors (mean and standard devia-

tion) computed 200 times in different positions using

ﬁve different ﬁshes.

Figure 8: Volume behavior after scanning a 1880 gr ﬁsh 200

times.

Table 1: Computed statistical descriptors of the volume of

ﬁve ﬁsh of different weights. Each ﬁsh was measured 200

times.

Fish

Figure

Mean

(mm

)

Std. Dev.

(mm

)

Max

(mm

)

Min

(mm

)

a 2425,6 5,87 2436,6 2406,4

b 2059,7 10,67 2080,5 2031,8

c 740,76 7,02 751,5 724,37

d 590,46 2,23 601,6 589,42

e 555,34 3,47 561,1 546,65

In the Table 1, It can be observed that computed

volumes have low variability. Only in the case of ﬁsh

2 the standard deviation (Std.Dev.) was superior to

10mm

. Another observed result is the fact that high

volume variability is not related to the ﬁsh volume.

The use of the density calculated in the Equa-

tion 12 and its use in the estimation of the voxel

weight (V

) is validated. For this, we take into ac-

count the customer restrictions. Speciﬁcally, the fact

that the weight deviation should not exceed the ±10gr

range, when it is selected the ﬁxed weight strategy

(i.e. cut all slices using a ﬁxed / previously selected

weight).

To test if this speciﬁcation is respected, we apply

the following steps:

• Scan one selected ﬁsh to determine each V

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

712

Table 2: The behavior of the weight for an 803 gr ﬁsh cut in 5 peace. 150 measures by each ﬁsh.

Weight (gr) Mean (gr) Std. Dev. (gr) Max (gr) Min (gr) Max-Weight (gr) Weight-Min (gr)

135 138,25 0,71 139,29 136,51 4,29 -1,51

140 145,35 0,38 146.41 144,53 6,41 -4,53

135 139,31 1,12 141.18 135,47 6,18 -0,47

121 124,84 1,20 127.75 120,83 6,75 0,17

94 96,44 1,80 99.90 91,36 5,90 2,64

• Weight the ﬁsh (T

) and calculate the density us-

ing the Equation 12.

• Using the ρ and V

, calculate the W

(Equation 13)

that is the weight associated to the i

image (see

Figure 6).

• Cut the ﬁsh at random places and measure the

slice length and weight using a rule and a preci-

sion scale.

• Using the length of the ﬁsh slice, estimate the

weight accumulated by each slice weight of the

proﬁles (W

Results can be seen in Figure 9. It shows 5 slices ob-

tained from a ﬁsh with 803gr of weight. We have

calculated the individual weight of each slice and

scanned 150 times. The Table 2 shows the minimum

and maximum obtained weight measured for each in-

dividual slice. In the two last columns, it is observed

that the weight of each individual slice has values that

are less than the established limit (±10gr).

Figure 9: Estimated weight of one ﬁsh of 803gr cut in

135gr, 140gr, 135gr, 121gr and 94gr.

From the Table 2, it can be deduced that the esti-

mated weights most of the times are greater than the

real weight of the ﬁsh slice. This does not represent

a problem for the proposed algorithms, because they

take into account the waste produced by the cutting

saw. With this, after cutting the ﬁsh, the actual weight

of the obtained slices is slightly lower than the esti-

mated weight. The volume lost at the time of cutting

is always less than W

and will depend on the thick-

ness of the cutting saw used and the ﬁsh’s area at the

cutting position.

The waste produced by the cutting saw near to the

tail ﬁn is smallest than the wasted produced near to

the head bone. As a positive aspect, in all cases, even

without considering the waste produced by the cutting

saw, always the ﬁnal weights are in the desired range

of ±10gr - the customer accepted deviation in relation

to the desired weight.

Taking in consideration the needs of the client,

the ﬁnal system prototype has been adjusted to pro-

duce more than 70 slices by minute. For this purpose,

the developed system has been optimized to exploit

the power of multicore CPUs by leveraging the use

of a multi-thread architecture. One thread control the

cameras to acquires the ﬁsh proﬁles and building the

top and bottom images, and the rest (the free threads)

are used for algorithms processing. This approach

provides real-time processing, and is able to guaran-

tee the customer’s requirements concerning the num-

ber of ﬁshes slices / minute.

The total delay of the automated vision system

is the sum of the acquisition time and the process-

ing time (see Table 3). Each ﬁsh takes between half

second and a second to be processed. In huge ﬁsh,

sometimes, the individual processing time of a ﬁsh is

slightly bigger than one second. With this, it can be

concluded that the automated vision system can per-

form almost in real time system.

Table 3: Behavior of acquisition and processing times for

different sizes of ﬁsh in milliseconds (ms).

Weight

(gr)

Acquisition

Time (ms)

Processing

Time (ms)

Max. Min. Max. Min.

2260 821 717 294 170

1880 749 671 286 129

670 681 610 217 54

520 451 440 98 49

510 422 410 82 46

Automated Vision System for Cutting Fixed-weight or Fixed-length Frozen Fish Portions

713

5 CONCLUSIONS

The main contribution of this work is the develop-

ment of an automated real time computer vision sys-

tem able to: 1) capturing and building a precise 3D

representation of frozen ﬁshes; and 2) accelerating the

automated frozen ﬁsh cutting in the production lines

at industrial level. An innovative high performance

image processing method has been developed that al-

lows to automatically infer the optimal place (coordi-

nates) where to cut the ﬁsh portions, attaining a pre-

viously wanted / selected weight or length of the pro-

duced slices and thus reducing the waste during the

cutting process. While the dataset used is still small,

the experimental statistical results demonstrated that

the proposed system prototype is feasible to be imple-

mented and validated in real industry environments.

6 FUTURE WORKS

The future work aims to validate the developed com-

puter vision system in real production lines. Also, it

is planned to improve the developed system with the

ability to evaluate the freshness of the produced ﬁsh

portions by analyzing properties like color, shape and

texture.

ACKNOWLEDGMENT

This article is one of the results associated to the

project MaxCut4Fish: Research and Development of

an Intelligent System for Automatic and Optimized

Cutting of Frozen Fish (project n

017664), sup-

ported by the European Regional Development Fund

(ERDF), through the Competitiveness and Interna-

tionalization Operational Program (COMPETE 2020)

under the PORTUGAL 2020 Partnership Agreement.

REFERENCES

Adamczak, L., Chmiel, M., Florowski, T., Pietrzak, D.,

Witkowski, M., and Barczak, T. (2018). The use of

3d scanning to determine the weight of the chicken

breast. Computers and Electronics in Agriculture,

155:394 – 399.

Booman, A., Márquez, A., Parin, M. A., and Zugarramurdi,

A. (2010). Design and testing of a ﬁsh bone separator

machine. Journal of Food Engineering, 100(3):474 –

479.

Food and of the United Nations, A. O. (2016). The state of

world ﬁsheries and aquaculture (soﬁa).

Kelkar, S., Boushey, C. J., and Okos, M. (2015). A method

to determine the density of foods using x-ray imaging.

Journal of Food Engineering, 159:36–41.

Kelkar, S., Stella, S., Boushey, C., and Okos, M. (2011).

Developing novel 3d measurement techniques and

prediction method for food density determination.

Procedia Food Science, 1:483 – 491. 11th Interna-

tional Congress on Engineering and Food (ICEF11).

Mathiassen, J. R., Misimi, E., Bondø, M., Veliyulin, E., and

Østvik, S. O. (2011). Trends in application of imag-

ing technologies to inspection of ﬁsh and ﬁsh prod-

ucts. Trends in Food Science & Technology, 22(6):257

– 275.

Mortensen, A. K., Lisouski, P., and Ahrendt, P. (2016).

Weight prediction of broiler chickens using 3d com-

puter vision. Computers and Electronics in Agricul-

ture, 123:319–326.

Poonnoy, P. and Chum-in, T. (2012). Estimation of sushi

shrimp weight using image analysis technique and

non-linear regression models. Internation Conference

of Agricultural Engineering (CICGR-AGENC).

Soltani, M., Omid, M., and Alimardani, R. (2015). Egg vol-

ume prediction using machine vision technique based

on pappus theorem and artiﬁcial neural network. Jour-

nal of Food Science and Technology, 52(5):3065–

3071.

Technologies, L. (2018). Gocator 2100 series.

https://lmi3d.com/products/gocator/g2/2100-series.

Last accessed 13 November 2018.

Viazzi, S., Hoestenberghe, S. V., Goddeeris, B., and Berck-

mans, D. (2015). Automatic mass estimation of jade

perch scortum barcoo by computer vision. Aquacul-

tural Engineering, 64:42 – 48.

ICPRAM 2019 - 8th International Conference on Pattern Recognition Applications and Methods

714