Microsoft Word - 1murphy.docx


 CHEMICAL ENGINEERING TRANSACTIONS  
 

VOL. 58, 2017 

A publication of 

 
The Italian Association 

of Chemical Engineering 
Online at www.aidic.it/cet 

Guest Editors: Remigio Berruto, Pietro Catania, Mariangela Vallone
Copyright © 2017, AIDIC Servizi S.r.l. 
ISBN 978-88-95608-52-5; ISSN 2283-9216 

Performance Optimization of Neural-Network Based Colour 
Measurement Tools for Food Applications 

Paolo Bargea, Lorenzo Combab, Paolo Gaya, Davide Ricauda Aimoninoa, Cristina 
Tortia*a, Nahid Aghilinateghc, Mohammad Jafar Dalvandc 
a
DI.S.A.F.A. – Università degli Studi di Torino, 2 Largo Paolo Braccini, 10095 Grugliasco (TO), Italy 

b
DENERG – Politecnico di Torino, 24 Corso Duca degli Abruzzi, 10129 Torino, Italy 

c
D.A.M.E. - University of Tehran, Karaj, Iran 

cristina.tortia@unito.it 

 
Colour is the first attribute subject to consumer perception in determining food quality and, in many cases, this 
is the only possible mean to qualify product at purchase. For this reason, the description of colour by analytical 
methods is fundamental in food processing control. 
Computer vision systems acquire RGB data which are device-dependent and sensitive to the different 
lightning. Therefore, they are not directly useful for colour evaluation to mimic human vision. On the contrary, 
traditional colorimeters, which adopt CIELab coordinates, work in human-oriented colour space where 
euclidean distance between two different colours (∆E) is well related to the difference perceived by human 
sight.  
Nevertheless, vision systems have many advantages as the capability of acquiring larger areas of the food 
surface and the easiness of implementation in automated plants at low costs.  
Neural networks, trained on a set of selected colour samples, can approximate RGB to L*a*b* relationships to 
characterise the colour of food samples under test.  
The aim of this paper is to present a rapid method based on neural networks for the calibration of a CCD 
(Charge-Coupled Device) camera colour acquisition system to obtain reliable L*a* b* information. Preliminary 
results concerning the influence of the composition of the training test and the camera settings (aperture and 
time of exposure) on the reliability and accuracy of the colour measurement system are also discussed.  

1. Introduction 

In food processing chains, colour is often checked on production lines to verify the product quality. 
Traditionally, colour inspection is performed by human eye or by colorimeters, which usually measure the 
CIELab coordinates with the consequence that the quality control is highly expensive in terms of manpower 
employed, the colour is affected by subjectivity or, as in the case of colorimeters, a very small area is 
considered. 
The colour acquisition of RGB values by camera sensors has the disadvantage to be dependent on the 
electronic sensor and on the illuminant. Methods to convert RGB encoded colours in the CIELab colour space 
have been proposed by different authors (Leon et al., 2006; Valous et al., 2009; Kılıç et Al., 2007; Taghadomi-
Saberi et al., 2015). Among these, Artificial Neural Networks gave good results.  
This paper aims to investigate the effect of the following factors on the performances of Artificial Neural 
Networks (ANN) colour calibration:  

1) number of hidden neurons in the ANN 
2) camera aperture and shutter 
3) colours used for training. 

 
The performance was tested separately for different colours using a commercial palette used by 
photographers for camera calibration. 

                               
 
 

 

 
   

                                                  
DOI: 10.3303/CET1758099

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Please cite this article as: Barge P., Comba L., Gay P., Ricauda D., Tortia C., Aghilinateg N., Dalvand M.J., 2017, Performance optimization of 
neural-network based colour measurement tools for food applications, Chemical Engineering Transactions, 58, 589-594   
DOI: 10.3303/CET1758099 

589



2. Methods 

A Nikon D7000 colour DSLR (Digital Single-Lens Reflex) camera with a CMOS (Complementary Metal-Oxyde 
Semiconductor) image sensor was mounted on a stand and connected with a dome lighting system, which 
ensured a uniform illumination of food samples. A white plastic hemisphere (350 mm diameter) reflected the 
light provided by two LEDs circular concentric crowns at the edge of the dome, obtaining a 5500 K colour 
temperature (Figure 1). The electric power of the lighting system was 8.45 W. The camera and the illumination 
system were placed into a (500 mm × 600 mm × 900 mm) wooden box internally black-painted to minimise 
external light and reflection (Valous et al., 2009). This system is a improvement of the equipment used in a 
previous work (Ricauda et al., 2015). 

 

Figure 1: Picture of the Computer Vision System 

 

Figure 2: X-rite Digital Colorchecker SG. Zone A and zone B are contoured in red (upper zone) and yellow 
(lower rectangle) respectively. 

The white balance was obtained using a X-rite white balance card. For each colour, 195 pictures of a 
ColorChecker Color Rendition chart were taken in all combinations of 11 apertures (5.6, 6.3, 7.1, 8, 9, 10, 11 
13, 14, 16, 18, 20 and 22) and time of exposure (1/15 s, 1/20 s, 1/25 s, 1/30 s, 1/40 s, 1/50 s, 1/60 s, 1/80 s, 

590



1/100 s, 1/125 s, 1/160 s, 1/200 s, 1/250 s, 1/320 s and 1/400 s). Camera ISO sensitivity was set to 200 and 
the image resolution at 300 x 300 dpi. Shooting sequence, shutter and aperture were automatically controlled 
by a PC running the software Digicam Control and a customised Matlab® program.  
The ColorChecker Color Rendition chart contains 24 patch of different colours (McCamy et al., 1976) and is 
widely used in photographic and video sectors for calibration purposes of imaging devices. Another 
ColorChecker version (Digital SG) was added with other colours (e.g. similar to skin, sky and grass) and more 
repetitions of grey-scale colours for a total of 140 patches. The manufacturer provides CIE L*a*b* coordinates 
for illuminant D50 (2 degree obsever) of the various patches for the two different ColorChecker versions as, 
for toxicity of pigments, in 2014 the firm had to change pigment formulations. Pictures of two zones of the two 
versions (February 2011 and July 2016) of the Munsell Digital ColorChecker SG (X-rite) were taken 
separately: a 6x4 patches zone containing the colours of the classic ColorChecker, which is the rectangle 
defined by the letters E2 (left upper corner) and J5 (right bottom corner) was defined as set A, and a set B, 
which refers to the zone defined by the E6 (left upper corner) and J9 (right bottom corner) patches (Figure 2). 
RGB values were obtained post-processing the acquired images by sampling RGB values in Matlab

®. From 
each colour patch 200 random 10x10 pixel square zones were sampled. Mean RGB values were used to train 
the neural network, for a total of 4800 RGB data used as input for each photo.  
CIE L*a*b* values (2° standard observer, D65 illuminant) were acquired by a Konica Minolta Chroma Meter 
CR-400 with a CR-A33f glass light-protection tube on Colorchecker. Five repetitions for each measurement 
were performed. 
The Artificial Neural Network (ANN) was composed by an input layer (Ni) containing the three RGB 
coordinates (Figure 3), a hidden layer (Nh) composed by a variable number of n neurons (with 7<n<12) and an 
output layer (No) which, in turn, contains the calculated values (O) for the CIELab coordinate considered (O=L, 
a or b). The feedforward Levenberg-Maruardt backpropagation method was used and run in the software 
Matlab® NN Toolbox (R2015b). 
 

 
 
Figure 3: Architecture of the neural network. Ni : input layer, Nh : hidden layer, No : output layer. 
 
Initial weights as well as bias values were fixed and taken by a randomly generated matrix in the -0.5÷0.5 
range. 
Input parameters were the three normalised RGB values while the output were the L* a* b* values. A neural 
network was obtained for each coordinate. The outputs of the three ANNs were then used to calculate the 
euclidean distance between the colours acquired by the images and the colorimeter data (indicated by 
circumflexed letters) following Eq. (1). 
 = ( ∗ − ∗) + ( ∗ − ∗) + ( ∗ − ∗) 	                 (1) 
 
For the ANN training, 58.8% of data were used for training while 23.6% for test, 17.6% for validation. Data 
acquired by colorimeter were used as target. 
A total of three ANNs (one for each of the CIELab coordinate) for each of the 195 pictures containing each 24 
colour patches were obtained. 
A matrix containing the different factors (shutter time, aperture, number of neurons of the ANN, photo, type of 
training) was analysed and the euclidean distance was considered as dependant variable. 

R

G

B

Nh

NoNi

O

n1

n2

n3

nn

n4

n5

591



3. Results 

The 2016 Colorchecker zone A colours were used to obtain the first series of ANNs. The ouputs of the ANNs 
were compared to the Cie L*a*b* values taken on the same zone of the 2011 Colorchecker. Similarly, the 
training of the network was performed on images of Colorchecker 2016 zone B and tested on the same set of 
the 2011 Colorchecker.  
A colour difference lower than 1.5 among the L*a*b* output of the ANNs and the measurements by colorimeter 
was considered as a detection threshold of a colour difference difficult to detect by the human eye (Guiné et 
al., 2015; Pathare et al., 2013). 
The analysis was performed on the whole matrix of 56160 cases: 195 images, 24 colours, six type of ANN (7, 
8, 9, 10, 11 and 12 neurons) and two colorchecker zone (A and B). 
The mean colour difference (Table 1) was quite high but it should be noted that saturated and underexposed 
images were also considered. 

Table 1:  Euclidean distance of the ANNs and colorimeter colour values on the two zone of the 2011 Color-
Checker. Mean colour difference and number of cases of patch colour  at ∆ 1.5 are reported. 

ANN neurons number Colorchecker zone A Colorchecker zone B 
 Mean ∆  % cases ∆ <1.5 Mean ∆ % cases ∆ <1.5
7 4.67 21.6% 6.25 9.9% 
8 4.07 27.1% 5.63 13.2% 
9 4.79 29.3% 5.58 14.9% 

10 3.71 32.5% 5.68 17.6% 
11 3.76 34.4% 5.24 19.7% 
12 3.29 35.9% 5.06 20.8% 

 
It was demonstrated that increasing the neurons number from 7 to 12 the ANN efficiency improved. This can 
be seen both from the mean euclidean distance as well as the rate of cases where differences among colours 
acquired by colorimeter and those calculated by the ANN have ∆ <1.5 which is considered a colour difference 
hard to be perceived by human eyes (Pathare et al., 2013). 
The colour difference was greater when the net was trained in zone B. This may be due to the fact that the 
colour palette was much more uniform. 
Some colours were poorly identified by the ANN, as shown in Figure 4. In many factors combinations, the 
ANN has reached the target to correctly recognise the colour (at < 1.5	∆ ) for colour present in both of two 
zones of the Color-Checker (Figure 4 and 5). 
 

 
 
Figure 4: Comparison of the ability of the ANN (12 neurons) to distinguish the different colours for both the 
zone of the 2011 Color-Checker. The number of photos by which colours are distinguished at euclidean 
distance above 1.5 are reported for each colour (Zone A colours). 
 

592



 

Figure 5: Comparison of the ability of the ANN (12 neurons) to distinguish the different colours for both the 
zone of the 2011 Color-Checker. The number of photos by which colours are distinguished at euclidean 
distance above 1.5 are reported for each colour (Zone B Colours). 

Nevertheless for some colours (E2, J2, H3, I3, E4, G4, H4 in zone A and J9, I9, H9, E9, I8, H8, G8, F8, I7, H7, 
G7 in zone B) the euclidean distance was very high. Among these, dark grey and black were also poorly 
discernible. 
It was observed that, using this method and choosing the 12 neurons ANN, it was possible to have good 
results only on maximum of 13 colours per photo, demonstrating that optimal shutter and exposure is different 
for each of the considered colours. 
For a possible use in food quality control the aperture/shutter combination for colour calibration of the ANNs 
should be chosen targeted to the dominant colour of the sample. In any case, the shooting of more than one 
image in different shutter/aperture combination could be envisaged. 

4. Conclusions and perspectives 

An Artificial Neural Network model was proposed for colour measurement which could be applied in food 
quality control. This tool has the advantage to consider a wider area respect to traditional colorimeters and to 
reduce costs for products colour checking. The method could be used for different food products and 
eventually paired to other on-line applications for image processing such as defect detection, classification 
and sorting. The proposed vision system could be adopted by food industries to control products during the 
processing phase, evaluating in advance possible defects and ensuring standard sensorial characteristics.  
Nevertheless, the system has proven to give good results only for some of the considered colours. For this 
reason, the improvement of the ANN learning is envisaged (e.g. increasing the number of colours in the 
training phase). 

Notation  

ANN Artificial Neural Network 
CIE Commission internationale de l'éclairage 
L* Luminance or Lightness component of colour in the CIELab colour space calculated by the neural 

network 
a* Red and green component of colour in the CIELab colour space calculated by the neural network 
b* Yellow/blue component of colour in the CIELab colour space calculated by the neural network ∗ Luminance or Lightness component of colour in the CIELab colour space measured by the 

colorimeter ∗ Red and green component of colour in the CIELab colour space  measured by the colorimeter ∗ Yellow/blue component of colour in the CIELab colour space  measured by the colorimeter ∆  Euclidean distance between two different colours 
CCD Charge-Coupled Device 

CMOS Complementary Metal-Oxyde Semiconductor 
RGB Red Green Blue  
DSLR Digital Single-Lens Reflex 

593



Acknowledgments  

This work was partially supported by the grants of the projects FoodVision and ColorEvolution (POR-FESR – 
Regione Piemonte).  

Reference  

Guiné R.P.F., Almeida C.F.F., Correia P.M.R., Mendes M., 2015. Modelling the influence of origin, packing 
and storage on water activity, colour and texture of almonds, hazelnuts and walnuts using Artificial Neural 
Networks, Food and Bioprocess Technology, 8, 1113-1125, doi 10.1007/s11947-015-1474-3 

Kılıç, K., Onal-Ulusoy, B., Yıldırımİsmail, M., Boyacı, H., 2007 Scanner-based color measurement in L*a*b* 
format with artificial neural networks (ANN), European Food Research & Technology, 226 (1), 121-126. 

Leon K., Mery D., Pedreschi F., Leon J., 2006. Color measurement in L*a*b* units from RGB digital images. 
Food research international, 39, 1084-1091, doi http://dx.doi.org/10.1016/j.foodres.2006.03.006. 

McCamy, C.S., Marcus, H., Davidson, J. G., 1976. A Color-Rendition Chart. Journal of Applied Photographic 
Engineering, 2(3), 95–99. 

Pathare, P.B., Opara, U. L., Al-Said F. A. J., 2013. Colour measurement and analysis in fresh and processed 
food: a review. Food and Bioprocess Technology, 6 (1), 36-60, doi 10.1007/s11947-012-0867-9 

Ricauda Aimonino, D., Barge, P., Comba, L., Gay, P., Occelli,  A., Tortia, C., 2015, Computer vision for 
laboratory quality control on frozen fruit, Chemical Engineering Transactions, 44, 175-180, 
doi:10.3303/CET1544030.  

Taghadomi-Saberi, S., Omid, M., Emam-Djomeh, Z., Faraji-Mahyari, K., 2015. Determination of Cherry Color 
Parameters during Ripening by Artificial Neural Network Assisted Image Processing Technique, J. Agr. 
Sci. Tech., 17, 589-600. 

Valous, N.A., Mendoza F., Sun, D.W., Allen, P., 2009. Colour calibration of a laboratory computer vision 
system for quality evaluation of pre-sliced hams, Meat Science, 81, 132-141. 

594