Proceedings of the International Conference on Man-Machine Systems (ICoMMS)

11 13 October 2009, Batu Ferringhi, Penang, MALAYSIA

Color Recognition Algorithm using a Neural Network

Model in Determining the Ripeness of a Banana
Paulraj M P1 Hema C.R1 R.Pranesh Krishnan1 Siti Sofiah Mohd Radzi2
1
School of Mechatronic Engineering, Universiti Malaysia Perlis (UniMAP), Perlis, Malaysia.
2
Electronics and Electrical Engineering with Communication, University of Sheffield, England.

Abstract- This paper presents a simple color recognition License Plate Recognition have proposed the application of
algorithm using a Neural Network model and applied to the information fusion enhanced color recognition technology
determine the ripeness of a banana. The captured image of the in stolen vehicle identification. [4]
banana is resized and its RGB color components are extracted. This application also can be useful in reducing time
The color components of the resized images are rescaled using a
simple heuristic method. Further, a histogram for the rescaled
consuming and work efficiency for customers and cashiers in
image is obtained and used as a feature vector to identify the billing counters. For instance this can be applied when buying
ripeness of the banana. A simple graphical user interface system fresh vegetables or fruits. Normally, the weighing system is
is developed in MATLAB that classifies the ripeness of the used for cashiers to determine the prices of the fruits or
banana. The proposed model has an accuracy of 96%. vegetables. However this color recognition system can
substitute the weighing system and it is capable to recognize
Keywords - Color Recognition, Neural Network, RGB, the type of fruits or vegetables based on the color of the
Histogram, Ripeness of Banana. scanned image.
Therefore, in this research, a simple algorithm method using
I. INTRODUCTION
the MATLAB is proposed to determine the ripeness color of a
fruit and the sample of the fruit used is banana. The image of
The normal human eyes have three types of sensors banana is captured and resized. Later, the image is extracted
and the signal of these three sensors determine the color into the RGB color components and each pixel of the color
response of the observer. The response of this system component is rescaled using a simple heuristic method. As a
produces the three-dimensional phenomenon of three result, histograms are obtained and used as the feature vector
dimensional spaces. When a human sees something, the light in determining the ripeness of the banana. A simple algorithm
enters the eye and hit the light detector on the retina. This method is proposed and the program is eventually presented
behaves similarly to a digital camera that records more using the graphical user interface.
reading whenever more lights hit the light detector on the back II. SYSTEM DESIGN
of the camera. Therefore in the electronics graphical system
RGB color model is similarly applied to provide such The RGB images of bananas are captured using a Web
coordinate system of three dimensional spaces. [1] Digital Camera. The captured images have the resolution of
In electronics graphic system, RGB color model is 320x240 pixels. In this research work, many environmental
commonly used for input and output devices such as color TV, have to be considered such as the position of the camera,
digital cameras, video recorder, computer screen, and LCD. lighting sensitivity, and background conditions as this
Computer monitors for instance use the RGB color and the elements have probabilities to effect the collection of the data.
computer images are commonly stored in RGB color system. The camera is placed 15 centimeters above the sample as in
[2] RGB color system is based on three basic color (A) in Figure 1. The banana is placed on a black paper that
components of red, green and blue in pixel {r, g, b}.In a true laid on the ground as in (C) in Figure 1. The device is
image color, for each color component, there is a range of equipped with two similar power of bulbs throughout the
intensity, from 0 to 255 and when these three combinations of capturing process in providing sufficient lighting to the sample
various intensities are superimposed, a colorized hue is as in (B) shown in Figure1. It is important to have similar and
obtained. [3] constant camera and light sources as different cameras and
From this theory, the color recognition model has been different light sources create different pixel intensities. [5]
applied widely in industrial sectors, commercial fields as well
as in social responsibilities. For instance, it is used as a III. METHODOLOGY
powerful and reliable parameter in robotics machines, aid for
the blind and the color blind people, diamond color sorting, This system is designed to identify the ripeness of the
quality control for the manufacture colored paper, [9] and in banana. There are four sets of bananas used in the research
characterizing the thermal paints. [10]. Rami Al-Hmouz and with different type of sizes and ripeness. Each set contains of
Subhash Challa in their research in the field of Automatic between 10-14 pieces of banana. Each image of the banana is

2B7-1
 Proceedings of the International Conference on Man-Machine Systems (ICoMMS)
11 13 October 2009, Batu Ferringhi, Penang, MALAYSIA

captured in four different positions and the images are that lies between 86-170, it will be rescaled to 122 and for
captured daily until all bananas turn to be rotten. It has three 171-255, the color component will be rescaled to 255.
main stages; preprocessing, feature extraction and ripeness Therefore color histograms are obtained for each color
classification. component by counting the number of pixels that have same
color scales in the image array as shown in the example in
Figure 3.5, 3.6 and 3.7.These histograms are used as the

Capturing Images

Resizing the images

Preprocessing
Extracting the RGB
components of the image

Rescaling the image

using simple heuristic
Feature Extraction method

Histograms are obtained

as feature vector

Neural Network
Ripeness
Classification
Fig. 1. Stand that holds the web digital camera at (A), two bulbs at (B) and the
banana will be laid on the ground on (C) Classified Ripeness Color

A. Preprocessing Fig. 2. System Block Diagram

In the preprocessing stages, the captured images are resized
in order to reduce the color index of the images. Each image feature vector to determine the ripeness of the banana by using
of the banana will be extracted by pixel into its red, green and a simple heuristic method.
blue color component as shown in Figure 3.2, 3.3 and 3.4. 100 samples of ripe banana will be allocated in the ripe
These red, green and blue color components are displayed in folder and 116 samples of the unripe banana will be
gray-scale index allocated in unripe folder. The results of these two types of
these samples are saved in an Excel file and the data will be
B. Feature Extraction sort out randomly and will used to train the network as the
In comparison with the mechanism of normal human eyes, a input patterns for the Neural Network developed sing
ripeness of the banana is determined by a simple glance of the MATLAB.
eyes to the majority color that dominates the banana. In a
micro scale observation, the color of a ripe banana is not C. Ripeness Classification
totally all yellow as there will some black dotted colors or still Artificial Neural Network (ANN) is a new form of artificial
any green-yellow color skin. Therefore, from these principles, intelligence of brain to compute and analyze the data. [7] In
each pixel of the color component of the image is analyzed training the Neural Network system, two methods are
and rescaled into three groups. normally used, the supervised and unsupervised methods. The
In the feature color extraction, the histograms are normally supervised method means that the system needs to be trained
used as a feature vector. For example, Swain and Ballard have by an instructor and if it is compared to the way of the brain,
introduced the original idea of using histograms in indexing of it means that the neurons act differently compared to the rest
large image database.[5].Besides, a binary histogram have in guiding the process. [8] For the unsupervised method, the
been proposed in utilizing the image gray scale level. In 1995, system learns by its own self to determine the process.
a color histogram is used as feature in color recognition. [6] However in this research, only supervised method is used
As for the intensity of color component that lies between 0- using a developed Neural Network model using the error back
85,the image will be rescaled to 0,for the color component propagation model.

2B7-2
 Proceedings of the International Conference on Man-Machine Systems (ICoMMS)
11 13 October 2009, Batu Ferringhi, Penang, MALAYSIA

The Network Architecture has one output layer of 9 input chosen as 400. The initial weights for this Neural Network are
neurons, 45 hidden neurons and one output layer with one randomized between -0.5 and +0.5. A trial weight set consist
output neuron. Both the hidden and output layer neurons are of 20 sets of randomized weight samples are considered. The
activated by binary sigmoid activation function. If the output testing tolerance used is 0.2. 60% samples are used for
is 1 then the fruit is ripe and if it is closer to zero then it training and all 100% samples are used for testing.

Figure 3.1 Resized image of banana

Figure 3.5 The number of pixels versus the color scales for the red component

Figure 3.2 Red component image of the banana

Figure 3.3 Green component image of the banana

Figure 3.6 The number of pixels versus the color scales for the green
component

Figure 3.4 Blue component image of the banana

represents the fruit to be unripe. The training tolerance of the

network is fixed as 0.01 and the maximum number of epoch is

2B7-3
 Proceedings of the International Conference on Man-Machine Systems (ICoMMS)
11 13 October 2009, Batu Ferringhi, Penang, MALAYSIA

[8] Juan R.Rabual and Julin Dorado (2006), Artificial Neural Networks in
Real-Life Applications, IGI Publishing.
[9] M Stoksik, D T Nguyen, M Czemkowaki, A Neural Net Based Color
Recognition System, University Of Tasmania, Australia.
[10] Tristan Lalanne and Christine Lempereur (). Color Recognition with a
Camera, A Supervised Algorithm Recognition, Toulouse Cedex, France.

Figure 3.7 The number of pixels versus the color scales for the blue
component

IV. CONCLUSION

In this paper, we developed a system for the purpose of

recognizing the ripeness of fruit. The system has three main
stages; preprocessing, feature extraction and ripeness
classification. The work was accomplished by training a set of
input data. After preprocessing, the feature extraction stage in
applied to each resized captured images. The images are
rescaled using a simple heuristic method. Histograms of total
pixels for each rescaled color threshold in an image are
obtained and used as a feature vector to determine the ripeness
of the fruit. A Neural Network model is developed for color
recognition. From the experimental results, the simulations
show that the ripeness recognition rate is 96%. In future, this
method can be applied in increasing the effectiveness for
cashiers and customers in counters when determining the
prices of fruits or vegetables instead of the conventional
weighing method as this method is simpler and will save time

REFERENCES
[1] Peter Shirley (2005), Fundamentals of Computer Graphics, A K Peters,
Ltd. is a leading independent scientific technical publisher based in
Wellesley, Massachusetts.
[2] Fortner, B. and Meyer, T. E. (1997), Number by Colors, Springer-
Verlag, New York.
[3] Sun, W. and He, Y. (1998), Spatial-chromatic clustering for color
image compression, Proceedings of I998 IEEE World Congress on
Computation Intelligence, FUZZ-IEEE, pp. 160 1 1604
[4] Rami Al-Hmouz, Subhash Challa (2007), Intelligent Stolen Vehicle
Detection using Video Sensing, Networked Sensors Technologies Lab,
University Of Technology Sidney
[5] Swain,M.J, and Ballard, D.H.,Indexing Via Color Histograms, In
Proceeding of Third International Conference on Computer Vision,pp.
390-393,1990
[6] Ennesser, F., and Medioni G. Finding Waldo, or Focus of Attention
Using Local Color Information, In IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol 17, No 8, pp. 805-809, Aug
1995,
[7] S.N. Sivanandam, M. Paulraj (2003), An Introduction to Neural
Networks, Vikhas Publications Company Ltd. India.

2B7-4