INTRODUCTION
Sugarcane (Saccharum sp.) is a clonally propagated grass
of the Gramineae family characterized by a high degree of polyploidy and
is a crop of major importance providing about 65% of the world sugar.
Reproductive tissue is harvested as the economic product in nearly all
field crops but this is not the case in sugarcane. In sugarcane, the stalks
are the harvested tissue and stalk size has a major influence on yield.
There has been virtually some research reported on the variation in size
of individual stalk internodes with position on the stalk and with crop
growth (Zambrano et al., 2003; Sinclair et al., 2005).
Sugarcane planting with traditional methods is costly, time-consuming
and necessary compression of buds in the field is not achieved easily
because of stalk planting in sugarcane. In tradition planting method,
great human force and high volume of sugarcane stalk in hectare is required.
To solve this problem and mechanizing of sugarcane planting, we suggest
the application of machine vision system and image processing methods
to identify nodes from sugarcane and to plant it as a seed by planting
machines. Many studies have been done in image processing method for contaminant
removal from wool (Zhang et al., 2005), discrimination of hard-to-pop
popcorn kernels (Yang et al., 2005), measurement of hot formed
parts (Dworkin and Nye, 2006), weed control system for tomatoes (Lee et
al., 1999), lentils grading (Shahin and Symons, 2001) and sorting
of apple (Shahin et al., 2002). Edge detection is an essential
technique in many image processing applications such as object recognition
and motion analysis. From the view of accuracy, this technique can be
classified into pixel-level and subpixel-level edge detection. Early edge
detection method employed local operators to locate edge with approximately
computing the first derivative or second derivative of the image gray
level step in the spatial domain, Prewitt, Sobel, Marr-Hildreth and Canny
operator are examples of pixel-level edge detection methods. Short running
time is the advantage of pixel-level edge detection method (Dong et
al., 2006).
This research was done to design and evaluate an identifying algorithm
for sugarcane nodes by using image processing and machine vision system.
In order to design an image processing system one should select a reasonable
algorithm in first and then apply the required hardware.
MATERIALS AND METHODS
For capturing image from sugarcane stalks during time intervals
in proportion to carrier belt speed, a CCD camera (CCD-TVR128E, Sony,
Japan) was used. These images were digitized through the use of image
card and stored as a two-dimensional array. Each value in the array had
an integer representing the light intensity of the corresponding pixel
(picture element). Monochrome images are stored as eight-bit integer values
in each array position while color images were typically stored using
24 bits per pixel, with eight bits representing the intensity of each
of red, green and blue light components (Khalili, 2000; Dworkin and Nye,
2006). A typical digital imaging system is shown in Fig.
1. The processing power of computer used in this research was 2/4
GHz. The related Digital data were processed by using algorithm and the
location of nodes was recognized. The algorithm used in this study for
identification of nodes was shown in Fig. 2. This study
was conducted in triplicate in the Center of Science and Research of the
University of Islamic Azad University of Tehran during 2006-2007.
|
| Fig. 2: |
Identification algorithm for sugarcane nodes |
RESULTS AND DISCUSSION
Normalization and Convolution of Image Pixel Values: Each color in RGB system had maximum light intensity of 0-225 for
8 bit which was obtained using Eq. 1 of RGB amounts for
every pixel (Eslami and Nazemi, 1998).
In fist stage, the resulted data form Eq. 1 was normalized
to increase calculation rate according to Eq. 2 (Nakajima
et al., 2003).
Normalized values of pixel (NVP) were about 0-225 which could be included
in a byte (8 bit). Normalization had no adverse effect on the identification
of sugarcane nods (Fig. 3a). In second stage, convolution
operation was used for identification of sharp edges. This operation combined
the amount of pixels in order to obtain desirable results (Eslami and
Nazemi, 1998; Kurita et al., 1998; Riesenhuber and Poggio, 1999).
Convolution operation was done using the Eq. 3.
where, C(X, Y) is pixel value after convolution, P(X, Y) is original
pixel value and M(X, Y) is convolution mask matrix. Different characteristics
of images can be obtained with changing of M(X, Y) matrix. The convolution
image was shown in Fig. 3b.
Eliminating Noise from Original Image Pixels: Threshold function was used for eliminating noise and reveal node
according to Eq. 4, which was done as look-up table operations
on the image pixels (Eslami and Nazemi, 1998).
In which TF(X, Y) is threshold function and k = 60 that was obtained
by error and trail. Therefore, image pixels values were changed to 0 (for
background) and 225 (for nodes) and low light intensity pixels were eliminated.
The threshold image of sugarcane stalk was shown in Fig.
3c. Resulted images were changed to negative mode in order to further
elimination of probable noises using LUT function (5) (Eslami and Nazemi,
1998; Khalili, 2000).
where, NF (X ,Y) is negative function. In this case, image pixels values
were changed to 0 (for nodes) and 225 (for background). The negative image
of sugarcane stalk was shown in Fig. 3d.
| Fig. 3: |
(a) Primary image, (b) image after convolution operation,
(c) image after using threshold function and (d) negative image and
(e) location of cutting points |
Algorithm for Node Identification: Node identification algorithm firstly evaluates row by row image pixels
from left corner. With progress of this stage, the number of pixels which
have zero amounts in a row were calculated and divided by total pixels
in a row. If the resulted value was grater than 0.154, it could be considered
as node (Fig. 3d). Therefore, algorithm identified the
nodes by using the following comment:
If NNP (number of node pixels)/NTP (number of total pixels)> 0.154
then
| Node |
= |
true |
| Else |
|
|
| Node |
= |
False |
In the next stage, identification of cut-points in sugarcane stalk occurred
by algorithm using the Eq. 6, 7 and
8.
where, TS is one third of distance between two consecutive nodes, H(i)
is the location of node = i, H(i+1) is the location of node = i+1, FP
is location of the first cut-point, SP is location of second cut-point.
In Fig. 3e, the location of nodes, background and cut-point
were displayed by dark, white and red colors, respectively. As shown in
Fig. 4, a graphical user interface (GUI) was designed
for mentioned algorithm by visual basic programming (visual basic No.
6.0).
As it was shown in the left side of Fig. 4, the sent
image from camera will be shown real-time. With setting the time interval
between two images or processing speed using the scrollbar and pushing
start button, images will send out to above section in the right side
and image processing stages will occur. Finally cut-point locations between
two consecutive nodes will be sent to micro-controller. Designed GUI had
two manual processing and automatic processing units. In manual processing
unit,
 |
| Fig. 4: |
Graphical User Interface (GUI) for identification of
sugarcane node |
 |
| Fig. 5: |
Error percentage of right sobel mask matrix with others |
desirable image is selected by pushing Insert image button and will be
processed when start button pushed. In this way, location of cut-points
will be appeared in the position of cutting. In automatic processing,
at first, processing speed should be set by scrollbar and then start button
in right side should be pushed. In the both units, the pixels proportion
of nodes could be changeable.
Appropriate Mask Matrix for Convolution Operation: Standard matrixes were used for selecting suitable matrix mask on
100 sugarcane node samples in 39 images and right sobel, left sobel, east
perwitt, west perwitt and western south perwitt were evaluated. Only the
right sobel could be able to identify the node locations in dark stalks
and low resolution pictures. SPSS statistical analysis showed that the
average and standard errors were 2.08 and 0.30%, respectively (Table
1). Error percentage of right sobel was compared with other mask matrixes
in Fig. 5 and 6. The X axel with zero
error percentage indicates the real location of nodes in sugarcane stalks.
It is clear that the error percentage of right sobel mask matrix is the
lowest.
The right sobel matrix was assigned as the best mask matrix with the
variance and standard deviation of 8.82 and 2.97, respectively. In addition
to identifying all of sugarcane nodes by using the suggested algorithm,
only 97.92±0.3% of the real locations of nodes was determined.
This was due to
 |
| Fig. 6: |
Error percentage of right sobel mask matrix with others |
 |
| Fig. 7: |
(a-b) Before and after processing image of sugarcane
stalk with high light intensity, (c-d) Low light intensity |
 |
| Fig. 8: |
Histogram of image data after convolution |
| Table 1: |
Statistical analysis of percent error of node location
by standard masks |
 |
 |
| Fig. 9: |
Image data after threshold operation |
irregular geometrical shape of nodes. Results showed that light intensity
had not significant effect on the identification of node location in this
method (Fig. 7). As the histogram (Fig.
8) indicates, the variation of image data with values less than 60
is very high and this noise can be eliminated. Therefore, by using threshold
operations (Fig. 9) the value of pixels are change to
0 and 225.
CONCLUSION
In present study, a method was introduced for identification of
node location on the sugarcane stalk. This approach consists of six steps
of capturing, digitizing, normalization, convolution, threshold and negative
operation. All of sugarcane nodes were identified using the suggested
algorithm. The advantage of this image processing system was high processing
speed with a run time less than 0.500 sec and high precision with a relative
error of measurement no more than 2.08±0.30%. These results are
in agreement with that of obtained by Dong et al. (2006). Right
sobel edge correction mask matrix had minimum variance and standard deviation
in comparison with other masks; therefore, it was selected as the best
mask for identification of sugarcane nodes. The proportion of node pixels
to total pixels in a row depend on factors such as image wide, camera
height from sugarcane stalks and camera magnification. When camera height
was 1 m, the number of pixels in row was 70 pixels and camera magnification
was selected in minimum amount, the proportion of node pixels to total
pixels was obtained as 0.154.
ACKNOWLEDGMENT
The authors would like to express their thanks to Dr. A. Homayouni
from Department of Food Science and Technology, Faculty of Health and
Nutrition, Tabriz University of Medical Science for his assistance.
Subscribe Today
