2012 Sixth International Conference on Sensing Technology (ICST)

An Intelligent Reflective Colour Sensor System for Paper

and Textile Industries
P Anupama1, Sathees Kumar K V1 , Dr S Rominus Valsalam1 , G Harikrishnan 1 , and
V Muralidharan2
1
Control and Instrumentation Group, Centre for Development of Advanced Computing (CDAC-T) ,
Thiruvananthapuram, Kerala, India
2
Centre for Development of Advanced Computing (CDAC) , Hyderabad , Andhra Pradesh , India

Abstract - The world is simply less colorful, without the sense architectures. Therefore, low cost, compact, and fast color
of colour. In ordinary life, a difference in colour perception is sensors must be developed for many important industrial,
mostly inconsequential. However, in many industries, the agricultural, and energy-saving applications. The most useful
ability to sense colour precisely can be crucial. In this paper, sensors will also include built-in, self-organized quality
a novel reflective colour sensing system is presented for monitoring features. These features will further reduce waste,
process monitoring and control applications in paper and improve biological and chemical sensing, enhance machine
textile industries. The system is developed using a solid state vision systems, and even have extension to energy
RGB sensor and a smart signal processing algorithm conservation, and to the efficiency of remediation and
implemented on micro-controller architecture. A hybrid recycling efforts.
neural network comprising Self organizing mapping and Back There are two basic colour sensing systems: reflective and
propagation architecture is used for colour zone classification transmissive. In reflective sensing, the sensor detects the
and exact colour identification of papers. Demonstrator colour of light reflected from a surface, while in transmissive
applications and simulation results are discussed to highlight sensing, the sensor detects the colour of light from a light
the importance of sensor and accuracy in measurement. source.In a reflective colour sensing system, the colour of the
light reflected off a surface is a function of the colour of the
Keywords: Colour correction matrix, Hybrid Neural
surface and the spectral content (wavelengths) of the
Network, Microcontroller, Pattern Classification, RGB sensor
illuminating light source. In a reflective colour measurement
system, the spectral content of the illuminating light source
must be considered along with the target object’s absorption
1 Introduction and reflectivity characteristics. In a transmissive color sensing
Color is an essential part of the paper and textile system, a sensor detects the light color and intensity emitted
industries. It greatly influences sales volume and is a primary from a source, after passing through a medium. Transmissive
factor in the purchase of a product. A common objective in sensing is used to measure mediums such as liquids, gas,
color technology is to control and reproduce a color under a plastics or glass.
set of specified conditions. In the paper and textile industry,
effective color control and communication between designer, The intelligent colour sensor presented in this paper makes
dyer and retailer are critical to obtaining high product quality use of low cost sensing devices while exploiting a smart signal
and cost efficiency. However, many of the factors enabling processing, based on a two layer neural network strategy
color control within the textile supply chain are not optimum, implemented on an embedded microcontroller. The sensor
and therefore it is likely that the critical path from product module developed is able to discriminate between a high
concept to consumer is extended, with associated cost number of colours, grouping colours with light differences in
inefficiencies. totality into a unique code while maintaining a suitable
resolution in colours discrimination. Conventional colour
Colour sensors enable solutions for colour discrimination, measuring systems use tri stimulus colourimeters that consist
determination, and measurement in a wide variety of of three photodiodes each covered with an optical filter for
applications such as colour feedback control in solid-state red, green and blue (or the complementary colours), or use
illumination and RGB backlight systems, colour adjustment in spectral photometers that use optical diffraction techniques
printers, industrial process quality controls, portable medical and an array of photodiodes. The smart colour sensor
diagnostic systems, and paper and product handling presented here requires no extra optical filtering sensor
equipment. Real-time response is critical in these industries, in architecture.
order to optimize process control and thus reduce waste.
Unfortunately, existing color segmentation techniques are
based on slow, complex, and bulky signal processing
techniques, which cannot be implemented in compact

978-1-4673-2248-5/12/$31.00 ©2012 IEEE 481

2 Sensor Architecture
The architecture of colour sensing system is depicted in
Fig 1. The sensing head of the system has been realized
through a RGB sensor and a matrix of high luminance D65
LED as the artificial light source. The latter makes the system
independent on environmental luminance and guarantees
always the same operating conditions in terms of background
luminance to the RGB sensor. The light source illuminates a
sample, and the light reflected from the sample or transmitted
through it is detected by a RGB colour sensor that has Red,
Green and Blue detectors. The R G B Sensor makes use of
TCS230DB photodiode sensor of Texas Instruments and it
consists of an array of photo diodes (16 *4). These photo
diodes detects the reflected wave from the target colour to be
identified and outputs unique R, G and B values for each
colour in terms of its frequency. These analog frequency
values correspond to digital RGB values for the corresponding
colour. These values are digitized by the micro controller
(MSP 430) to get raw R, G, B values which need to be further
processed and corrected according to the spectrum of the
illuminant used, the spectral response of intervening optical Fig .1 Architecture of colour sensing system
components and the response curves of the sensor.
3 Device Characterization
Signals coming from the RGB sensors are then conveyed
through a dedicated electronics module for signal processing. In order to perform the device characterization a suitable
The data processing operation can be represented in matrix set of reference colour patterns was used. At any instant, for a
form as x = M.t .The symbol t represents a three-element particular process colour, the colour sensor chip produces an
vector containing the device values captured from any colour output in terms of RGB. All the existing colour sensors
patch. M represents the 3x3 colour correction matrix. The classify the colour spectrum into a set of 24 colour classes as
symbol x represents the resulting vector of estimated [R, G, B] pattern distribution along the RGB hyperspace is quite
values. Signals received from the sensor are acquired by the complex to be interpolated via a set of crisp rules. Due to this
microcontroller in the electronics module and elaborated to difficulty, classification and identification task is performed by
produce a unique identification of the class to which the target a 2-tier Neural Network. Colour classification job is
colour. In paper industries, CIELAB color space is used for accomplished by a hybrid neural network comprising a Self
precise measurement of colour. So the resulting RGB values Organizing Mapping Neural Network (SOM NN) for pattern
are converted to CIE L*a*b* colour space .First RGB values classification and a Multi- Layered Perceptron with algorithm
are converted to XYZ coordinates using the equations of Back Propagation Neural Network (BP NN) for exact
colour determination. The RGB values of colour samples are
X = R * 0.490 + G * 0.310 + B * 0.200 input as features of training samples and are classified by
Y = R * 0.177 + G * 0.813 + B * 0.011 SOM NN. Afterwards, the outputs of SOM NN are
Z = R * 0.000 + G * 0.010 + B * 0.990 respectively delivered to a set of BP NN for further learning.
This learning system converges fast, producing minimum error
Then XYZ are converted to CIELAB using the equations discrepancy.

CIE-L* = (116 * [(Y/Y0)1/3)]) – 16; A Self-Organizing Map (SOM) is an unsupervised and

1 ” 100Y ” 100 nonparametric network as proposed by Kohonen in early
CIE-a* = 500 * [(X/X0)1/3 - (Y/Y0)1/3)] 1980’s [7]. It converts patterns of arbitrary dimensionality into
CIE-b* = 200 * [(Y/Y0)1/3 - (Z/Z0)1/3)] the responses of two–dimensional arrays of neurons (feature
maps). One of the important characteristics of the SOM is that
A PC based interface has been realized for colour presentation the feature map preserves neighborhood relations of the input
with USB and serial (RS232 / RS485) communication pattern. A typical SOM structure consists of an input layer and
interface for process control. In-built 4*20 LCD display is also an output layer. The number of input neurons is equal to the
available for the remote operating of the device. A three dimensions of input data, and the output neurons are arranged
channel 4-20 mA analog output is also provided for process in a two-dimensional array. The connections between two
control applications. layers are fully connected.

482
 Let A denote set of output neurons and i to denote ith output wkj (t + 1) = wkj (t ) + ηδ pkOpj
neuron (i.e., i ⊂ A) .The input vector
w ji ( t + 1) = wji ( t ) + ηδ pjOp
x j = ª« x j1 , x ....x jp º» belongs to a p-dimensional Euclidean
¬ j2 ¼
space R ( x j ∈ R ; j = 1, 2...n )
p p The net is trained by presenting a set of samples (training set)
Each input vector is shown in Fig 2. Both of the utilized learning algorithms follow
connected to an output neuron through a reference vector the descending gradient of the error function E calculated on
output:
wi = ª wi1 , w ....wip º .Output neuron c is the winner for
¬ ¼ 1
E= ¦ (tpj − opj ) 2.
i2
input xj at time step t of learning if 2
|| x − wc ( t ) ||= min
j i∈ A
{|| x j − wi ( t ) ||} .Weight is Where tpj is the target corresponding to the pattern p for the
updated as output neuron j and opj is the output of the output neuron j.
the error appraised on output level is handed down to the
wm (t + 1) = wm (t ) + α [ x j − wm (t )]; m ∈ N c ( t ) neurons which are on the inferior levels.

wm ( t + 1) = wm ( t ) ; m ∉ N c ( t ) 4 Results
The main algorithm implemented in the microcontroller
Where Nc (t) represents the neighbors’ of neuron c at time t
device performs the normalization and the elaboration tasks.
and Į is a learning constant. The result of this learning process Colour codes estimated through the classification strategy is
is that the network tends to be organized with respect to the then conveyed to the PC interface via an RS-232/USB
input patterns. connection. Results indicate that the artificial neural network
was well able to recognize the given colour patterns. Matrix M
By way of the above learning structure, a non-linear model is used to correct the values obtained from the sensor so that
of colour identification is obtained. After colour samples are these values will be converging inside the required triangular
self organized and classified by the SOM NN network, data is region of the chromaticity diagram and therefore will be
categorized in clusters as a result of characteristic difference flexible enough to represent R G B values as shown in Fig 2.
thereof. The result of this learning process is that the network Matrix M is given by
tends to be organized with input patterns. After the SOM § 3.24 −1.53 −0.49 ·
M = ¨¨ 0.826 0.023 ¸¸
neural network converge to a balance state, the original image 1.752
is mapped from a high colour space to a smaller space. Then ¨ 0.0434 −0.1989 1.063 ¸
the data are respectively sent to BP NN for learning where by © ¹
the learning system quickly converges. The neural networks
used in this paper , back propagation neural networks was
chosen because of simplicity and its capability in supervised
pattern matching.

A fully connected multilayer network with Back-

Propagation is applied. For the activation function the
symmetric sigmoid function is used. Before training, the initial
weights are set to random values between -0.1 and 0.1.
Let be input applied. The values of
net pj for hidden layers are given by
.The output is computed
as .
Fig 2: Chromaticity diagram
The error term for output layer
The mean square error in the overall Back propagation is
is computed to
estimated. Results obtained during the experimental survey
determine the hidden layer revealed a high stability and repeatability of the device which
error, .The output and shows very poor data dispersion around averaged values. As
hidden layer weights are updated as can be observed a suitable efficiency of the recognition task
performed by the smart colour sensor is evincible, and the
network accurately determines the required colour with

483
minimum percentage error. Second test is conducted to verify
the repeatability and in this case also, good performance of
device is illustrated.

5 Demonstration Applications
We have demonstrated the developed reflective colour
sensor system in one of the leading paper plants of the
country. The results obtained were encouraging to take up
enhancement activities in the area of online reflective colour
sensing system. Fig 3. Shows the picture taken from the paper
industry where the system was installed . Fig 4. Shows the
colour sensor controller module and Fig 5(a) and Fig 5(b)
shows the reflective colour sensor module.

Fig.5(a). Reflective Colour Sensor module (Top View)

Reflective colour sensor module

Fig.3. Reflective colour sensor installed in paper plant

Fig.4. Colour Sensor Controller module

Fig.5(b). Reflective Colour Sensor module (Bottom View)

484
6 Conclusions [5] Y.C. Frankie and M.S Dmnokan,”Fiber optic sensor for
colour detection”, LEOS 1992, pp. 458-459, 1992.
A genuine smart reflective colour sensor has been realized [6] Sakurai, M. Kurihara and Y. Karasawa, S., “Colour
using two tier Neural Network. We have divided the problem classification using fuzzy inference and genetic algorithm”,
into four stages colour sensing, matrix correction, colour zone IEEE World Congress on Computational Intelligence, Third
classification and exact colour determination and worked out IEEE Conference on Fuzzy Systems, Vol. 3, pp. 1975-1978,
implementation strategy for each stage. The test results are 1994.
encouraging enough as discussed earlier. It is highlighted that [7] Tenvo Kohonen, “The self-organizing map”, Proc. of
the colour physics associated with the photodiode sensor array IEEE, Vol.78, No.90, pp1464-1480, Sept. 1990
and the neural network models are the right choice to address [8] Po-Tong Wang and Ching-Han Chen,”Hybrid Neural
the problem of accurate colour measurement in paper and Networks for Colour identification”, Patent U.S 6 571 228 B1
textile industries. ,2003
[9] H. A. Rowley, S. Baluja, and T. Kanade, "Neural network-
based face detection," in IEEE Transactionson on Pattern
7 References Analysis and Machine Intelligence, Vol. 20, pp. 23-38, 1998
[10] Hush, D.R.Horne, B.G, “Progress in Supervised Neural
Networks”, IEEE Signal Processing Magazine, Vol.10:1, pp.
8-39 1993
[1] Bruno Ando, Alberto Ascia, Salvatore Baglio, Nicola [11] QI Yingjian, LUO Siwei and LI Jianyu, “Using Neural
Pitrone”A smart colour sensor based on a Multi-Fuzzy Network in colour classification”, Proc.of TENCON '02 IEEE
Inference strategy”, Instrumentation and Measurement Region 10 Conference on Computers, Communications,
Technology Conference - IMTC 2007 Control and Power Engineering, 2002
[2] R.F. Wolffenbuttel. and G. de Graaf, “Performance of an [12] S. Haykin. “Neural Networks: A Comprehensive
Integrated Silicon colour sensor with a digital output in terms Foundation”. Prentice
of response to colours in the colour triangle”, Sensors and Hall PTR, Upper Saddle River, NJ, USA, 1998.
Actuators, A21-A23, pp. 574-580, 1990. [13] Emile Fieslera, Tuan Duongb, Alexander Trunova
[3] Chia-Chang Tong, Kun-Li Wen, Yu-Tusn Wang, Shuen- “Design of neural network-based microchip for color
Jeng Lin, Yao-Lun Liu and Chien-Ming Lee, “The segmentation ”
Development of Portable Infrared Colour Sensor”, IEEE Int.
Conference on Industrial Technology, pp. 959- 962, 2005.
[4] Klaus Eberhardt, Thomas Neidlinger, and Markus B.
Schubert, “Three-Colour Sensor Based on Amorphous n-i-P-i-
n Laver Sequence,” IEEE Trans. on Electr. Devices, Vol. 42,
N. 10, 1995.

485