Image Processing for Stress Cracks
in Corn Kernels
S. Gunasekaran, T. M. Cooper, A. G. Berlage, P. Krishnan
ASSOC. MEMBER
ASAE
AFFILIATE
ASAE
MEMBER
ASAE
ASSOC. MEMBER
ASAE
ABSTRACT
N image processing algorithm was developed for
detecting stress cracks in corn kernels using a
commercial vision system. White light in back-lighting
mode with black-coated background having a small
aperture for the light provided the best viewing
conditions. The kernel images, when processed using the
algorithm developed, produced white streaks
corresponding to the stress cracks. Double stress cracks
were the easiest to detect. Careful positioning of the
kernel over the lighting aperture was necessary for
satisfactory detection of single and multiple stress
cracks.
At present, corn kernels are examined for stress cracks
by candling the kernels with a bright light source from
below. This procedure is time consuming and fatiguing
to the human eye. Gunasekaran and Paulsen (1986)
investigated several nondestructive testing techniques for
stress crack evaluation. Light reflectance measurement
using a helium-neon laser light (632.8 nm) was found
insufficient for stress crack detection because reflectance
is a surface phenomenon (Gunasekaran et al., 1986).
The laser optical method, however, was very successful in
detecting external kernel defects. Ultrasonic imaging
was also found to be unsuitable because of the need for
slicing the kernel very thin to obtain a visual image using
ultrasonic energy (Gunasekaran et al., 1986). However,
optical imaging has been found to be suitable for stress
INTRODUCTION
crack evaluation (Gunasekaran, 1985; Gunasekaran and
Physical and mechanical stresses developed in corn Paulsen, 1986). Optical imaging or image processing is a
kernels as they are harvested, dried, stored, and handled relatively new technique that holds promise for
induce various quality defects. Defects such as surface- automatic, on-line quality evaluation and control of
splits, starch cracks and chip-offs, caused by mechanical wide-ranging materials (Gagliardi et al., 1984; Kahwati
stresses, are external and are easily detectable. However, and Law, 1985). This method essentially duplicates the
stress cracks, caused by a combination of thermal, condition as the eye sees an object. A typical image
moisture and mechanical stresses, are internal and not processing system receives light from a source; converts
readily identifiable.
the light into an electrical signal proportional to the
Stress cracks can be defined as very fine fissures in intensity of the light received; processes the analog
kernel endosperm u n d e r n e a t h the p e r i c a r p . electrical signals into a digital form usable by a
Gunasekaran et al. (1985) investigated the size computer; measures and analyzes various characteristics
characteristics of stress cracks with the help of scanning of the digital data representing the image; and interprets
electron microscope pictures. They concluded that a the image data to obtain useful information. The
typical stress crack is about 53 fjim in width and half the resolution of the digital image depends on the number of
kernel thickness in depth. The stress cracks were also pixels (picture elements) digitized for each scan line and
found to propagate from the center of the kernel and on the number of scan lines used. Proper lighting
may not extend to the surface underneath the pericarp. conditions are very important for processing speed and
In general, kernels with stress cracks have lower efficiency (Berlage et al., 1984; Gagliardi et al., 1984;
mechanical strength (Gunasekaran and Paulsen, 1985) Paulsen and McClure, 1985).
and break more readily upon subsequent handling than
Image processing applications in agricultural
non stress-cracked kernels. Amount of broken corn engineering are rapidly expanding. Quality evaluation of
along with foreign matter is a major corn grading factor. various biological materials such as apples (Graf et al.,
Cracked and broken corn yield lower percentages of 1981); brown rice (Matshuisa and Hosokowa, 1981); fish
large grits in dry milling and starch in wet milling than (Shimatachi et al., 1982); seed contaminants (Berlage et
undamaged corn. Large grits and starch are two of the al., 1984); and tomatoes (Sarkar and Wolfe, 1985) has
valuable prime products of dry and wet milling, been achieved by image processing. Image processing
respectively. Moreover, damaged corn is more readily has also been used as an aid to automatic fruit and
attacked by microorganisms; and thus has a lower vegetable harvesting (Whittaker et al., 1984; Sites and
storage life than undamaged corn.
Delwiche, 1985).
This article presents the application of image
processing technique for detecting stress cracks in corn
Article was submitted for publication in July, 1986; reviewed and
kernels.
approved for publication by the Electrical and Electronic Systems Div.
A
of ASAE in December, 1986.
Published as Miscellaneous Paper No. 1154 of the Delaware
Agricultural Experiment Station.
The authors are: S. GUNASEKARAN, Assistant Professor,
Agricultural Engineering Dept., University of Delaware, Newark; T.
M. COOPER, Senior Research Associate, A. G. BERLAGE, Research
Leader, USDA-ARS, Oregon State University, Corvallis; and P.
KRISHNAN, Assistant Professor, Agricultural Engineering Dept.,
University of Delaware, Newark.
266
OBJECTIVES
The objectives of this investigation were:
1. To determine optimal condition for viewing stress
cracks in corn kernels using a computer vision system.
2. To develop an image processing algorithm to
identify the presence of stress cracks in corn kernels.
© 1987 American Society of Agricultural Engineers 0001-2351/87/3001-0266$02.00
TRANSACTIONS of the ASAE
Printer
Disk
drives
Host
microcomputer
%
Vision system computer
Processing
module
X
Digitizer and
display module
Camera / monitor
interface
Video
camera
Black & white
monitor
Fig. 1—Block diagram of the vision system hardware.
SYSTEM DESCRIPTION
A commercial vision system, Intelledex V200*, was
acquired and subsequently upgraded. It was developed
with special hardware to interface with cameras and
display monitors and with software to implement
processing algorithms and drive a vision-to-robot
interface. Fig. 1 shows a block diagram of the system
hardware.
A Hitachi KP-120 solid state video camera was used
for image acquistion. A C-mount to bayonet adaptor
allowed use of 35-mm SLR photographic lens system.
The camera was mounted on a vertically adjustable stand
for necessary magnification and resolution. The stand
also provided support for lighting sources.
In the vision system, the analog camera image is sent
to the system computer. The computer consists of three
modules: (a) c a m e r a / m o n i t o r interface, (b)
digitizer/display module, and (c) processing module.
The camera/monitor interface module passes
information between the digitizer/display and the
camera and monitor. It also controls the gain, timing,
and selection of camera from which the image is
produced.
The digitizer/display module converts the analog
camera signal to a digital form that the computer can
process and store. Each of the 256 camera scan lines is
digitized into a series of 256 discrete picture elements
(pixels). Each pixel in this 256 x 256 array has a 6-bit
value (range: 0-63) representing the average light
intensity over its area. A value of zero is black while a
value of 63 is white. The hardware digitizes an image in
0.0167 s. The module has 64 k x 6 bits of static random
access memory (RAM) which is used for storing a single
digitized image. This image buffer or display RAM holds
a single frame for processing or display.
* Mention of trade names in this publication is solely for the purpose
of providing specific information and does not constitute endorsement
by the University of Delaware and the USD A over others of a similar
nature not mentioned.
Vol. 30(l):January-February, 1987
f
Fig. 2—Corn kernels with none, single, double and multiple stress
cracks from left to right, respectively.
The processing module contains an 8 MHz 8086
central processing unit (CPU), an 8087 numeric
coprocessor, 132 kbytes of read only memory (ROM) for
VISION BASIC, 256 kbytes of dynamic RAM, and 96
kbytes of CMOS battery-backed RAM. This module
executes the vision commands which control the
operation of the vision system. Vision generated data are
used in decision-making algorithms which are fixed stepby-step procedures for accomplishing a given
computational task.
The host microcomputer serves as an intelligent
terminal. Its main function is to execute the host
program which gives the vision system computer access
to the host microcomputer's disk drives, screen, and
keyboard. A 23-cm (9-in.) black-and-white monitor
displays the contents of the vision display RAM. This can
be a stored image, or a processed image.
PROCEDURE
Sample Selection
Corn kernels of four varieties namely, FRB73 x Mol7,
FRB73 X Va22, Mol7 x HlOO, and (FR4A x FR4C) x
Mol7 were used in this investigation. These corn
genotypes are grown popularly in the Midwest. The
kernels for the experiment were drawn at random from
high-temperature dried samples and grouped as those
with no, single, double, and multiple stress cracks (Fig.
2). About 25 kernels of each variety and of each stress
crack category were viewed individually under the vision
system to obtain the images.
Illumination
The need for proper lighting conditions for efficient
image processing has been well established. In this
investigation the kernels were illuminated using different
lighting modes such as front-lighting, back-lighting, and
side lighting.
Front-lighting, (illuminating from above the kernels)
and back-lighting (illuminating from below the kernel)
were provided using the Schott, Model KL1500 fiber
optic light source. This light source had a maximum of
150 W power rating and provided a maximum light
intensity of about 10 Mix at the fiber optic light guide. A
ring light guide mounted on the camera lens was used to
obtain a shadow free diffuse lighting for front-lighting.
For back-lighting, a flexible fiber optic bundle of 4.5 mm
diameter was used. Light intensities at the kernel
surface, when illuminated with white light, were
measured to be 48.3 klx and 32.3 klx under front and
back-lighting modes, respectively. A Wollensak Fastax
WF327 type light meter was used for light intensity
measurements.
The wavelength of the light used for illumination was
267
Fig. 3—Kernel viewing section of the vision system. 1 - Monitor; 2
-Camera; 3 - Fiber optic light source; 4 - Light guide for back-lighting;
5 - Light guide for front-lighting; 6 - Black-coated background.
varied using insert filters at the light source. While using
white light, the wavelength received by the camera lens
was controlled by mounting a suitable filter over the
camera lens. A series of filters namely, violet (Wratten
34A, 370 nm); blue (Wratten 47 B, 450 nm); Green
(Wratten 61, 515 nm); and red (Wratten 70, 610 nm),
were used. For side-lighting a pair of incandescent lamps
(18 W) mounted at 45° angle on either side of the kernel
were used.
The kernels were placed directly under the camera
over different backgrounds of frosted glass plate, milky
white glass plate, and black-coated wooden plate. These
backgrounds were used with each of the above lighting
modes to determine the best viewing conditions. The
black-coated wooden plate used with back-lighting had a
2.4 mm diameter opening for the light to pass through.
Fig. 3 shows a kernel being viewed with back-lighting
over the black-coated wooden plate background. This
figure also shows the ring light guide mounted on the
camera lens for front-lighting.
Image Processing
Image acquisition and processing were perfromed
using a variety of processing algorithms available with
the vision system. The algorithm used for stress crack
evaluation is similar to high-pass filtering. The pixels
representing stress cracks had significantly different gray
scale values than the pixels of the rest of the kernel
surface. Therefore, gray scale levels of the pixels
representing the stress cracks were extracted first by
creating an image supressing the gray scale levels of the
stress crack part and subtracting this newly created
image from the original image. The high-pass filtering
used in the image processing algorithm passes high
frequencies in the gray scale value, i.e., it passes the
pixels with large change in gray scale value (derivative) in
relation to the neighboring pixels but not necessarily
those pixels with high gray scale values. Following is the
sequence and brief description of each of the step:
VSNAP acquires the real-time digitized image of the
object under the camera and writes it into the display
RAM.
VDIG is a mode command and does not perform any
processing function. It switches the signal shown on the
monitor to the current image in display RAM. This step
is required only to see the actual processing of the image.
268
The VDIG display is static, and is not affected by the
object under the camera. Therefore, once VSNAP and
VDIG are performed the sample can be removed from
viewing position. This enables acquiring several images
at one time for later processing.
VSIMAGE stores the digitized image in an image
buffer. This original image is later used in the
VSUBTRACT step to extract the stress cracks.
VENHANCE performs an image enhancement
operation on the contents of display RAM. Each pixel in
the image is arithmetically averaged with its eight
surrounding pixels and a new pixel value is obtained.
The action is similar to low-pass filtering; and has the
effect of smoothing the contrast of the
pixels
respresenting the stress cracks that have significantly
high gray-scale value compared to their surrounding
pixels.
VSUBTRACT numerically subtracts the gray-scale
value of each pixel of the current VENHANCEd image
from the corresponding pixels of the original image
stored in the image buffer (obtained by VSIMAGE). As
mentioned above, this has the effect of passing those
pixels with large rate of change or derivative of gray scale
value. Change in gray scale value is usually large at
places of discontinuity like a crack.
VTHRESH operates based on the threshold gray-scale
value used (the numeral following the command). All the
pixels with gray-scale value less than the threshold are set
to the binary value of zero (black); the rest are set to one
(white). Thus, VTHRESH produces a purely black and
white image.
The images obtained using the above steps showed
white streaks corresponding to the presence of stress
cracks; but they also contained some spurious streaks
which could be mistaken for stress cracks. To eliminate
these spurious streaks the image-smoothing step,
VENHANCE was repeated several times before
VSUBTRACT operation. Additional VENHANCE
operations eliminated the spurious streaks, but it also
eroded the streaks representing stress cracks. After
several trials, repeating VENHANCE three times was
found to be optimal.
In order to obtain a better stress crack recognition, a
contrast enhancement algorithm (VMAP) was added to
the program prior to VENHANCE operation.
VMAP enables remapping of any or all pixel grayscale values to new values as specified. In the program
used, all pixels with gray-scale values less than a chosen
lower limit (XL) were set to zero. The pixels with grayscale values greater than (XL + 32) were set to 63. Those
pixels with gray-scale values from XL to (XL + 31) were
remapped as 2x(I-XL), where I is the gray-scale value of
the current pixel. This operation has the effect of
doubling the contrast of those pixels with gray-scale
values from XL to (XL + 31). The lower limit XL was
chosen to be 20 by examining the gray-scale histogram of
the original image.
The program with VMAP eliminated the spurious
streaks and spots within the kernel boundary more
effectively. However, this also caused the streaks
representing the stress cracks to split and shorten.
Therefore, an image-structuring algorithm VDILATE
was added to the program.
VCOMPRESS compresses the image in display RAM
into bit plane zero. A bit plane is a contiguous 8 kbyte
TRANSACTIONS of the ASAE
block of RAM located in a saved image buffer. The
vision system has bit planes 0 to 7 representing 64 k
RAM. VCOMPRESS is necessary for the VDILATE
operation.
VDILATE algorithm dilates contents of bit plane zero
by a linear structuring element. Dilation of an image by a
structuring element can be defined as the union of the
translations of all points in the image to all points in the
structuring element. The software allows use of 8
structuring elements, each in a different direction. For
stress crack detection, structuring elements in any two
opposite directions were found most suitable.
A BASIC program of the final version of the
processing algorithm is given in the Appendix.
RESULTS AND DISCUSSION
Front-lighting generally produced lower contrast
images than back-lighting. Therefore, even though the
stress cracks were visible on the monitor with frontlighting, the information was lost in the processing steps.
This is expected because front-lighting images are
similar to light reflectance measurements; and reveal
only the surface characteristics. Images obtained with
side-lighting were similar to those obtained with frontlighting. Even though side-lighting is expected to reveal
more of inner characterisitics than front-lighting because
of the body reflectance (Gunasekaran et al., 1986), it was
not sufficient to bring out the stress cracks. Backlighting, on the other hand, produced images with high
contrast between stress cracks and the rest of the kernel
surface. Thus, the images obtained with back-lighting,
when processed, exhibited bright streaks representing
the stress cracks. Back-lighting used in the experiments
is similar to the candling procedure currently in use for
stress crack detection. The frosted glass and milky white
plates used as backgrounds for placing the kernel did not
give good results because of huge light dispersion around
the kernel. The black-coated wooden plate with a
2.4-mm opening helped to eliminate the light dispersion
and provided a better contrast for stress cracks than
other two backgrounds.
Filters of varying wavelengths used both at the light
source and at the camera lens diminished the light
intensity reaching the lens and hence produced a lower
contrast image than with white light. Therefore, it was
determined that using white light (no filters) in the backlighting mode with black-coated plate having a small
aperture as the background to be the optimal lighting
and viewing condition.
The original images of the kernels with single, double,
multiple stress cracks and the corresponding processed
images are shown in Fig. 4, 5, and 6, respectively. In the
processed images the outer lines partially define the
kernel periphery; and the stress cracks are represented
by the lines inside the kernel boundary. The algorithm
performed very satisfactorily in extracting the stress
crack details in 90% of the kernels examined. The
success rate was determined by comparing the visual
evaluation of the kernels for stress cracks with the
corresponding evaluation of the vision system using the
same set of kernels. Of the three stress crack categories,
double stress cracks were the easiest to detect. This is
possible because most double stress cracks branch out on
either side of the kernel surface, which become highly
visible when light is directed at the center of the kernel.
Vol. 30(l):January-February, 1987
Fig. 4—Original and processed images of a single stress-cracked corn
kernel.
In case of single and multiple stress-cracked kernels, the
cracks were not clearly seen if the lighting was directly
below. Therefore, it was necessary to position the kernel
over the aperture slightly off-center to obtain the stress
crack images. This problem was more serious with single
stress cracks than with multiple stress cracks, because,
in case of multiple stress cracks, those parts of the cracks
that are around the center of the kernel (or around the
lighting area) were visible. One way of solving this
problem in an automatic system is by scanning the kernel
surface by changing the kernel position two or three
Fig. 5—Original and processed images of double stress-cracked corn
kernel.
269
centering the kernels over the lighting aperture.
5. Single and multiple stress-cracked kernels
required careful positioning over the lighting aperture in
order to obtain complete stress crack details.
6. Further improvement in the processing algorithm
and/or illumination condition is required to obtain
information about very fine cracks and complete crack
length.
References
Fig. 6—Original and processed images of a multiple stress-cracked
corn kernel.
times, to account for all of the stress cracks.
In all three stress crack categroies the faint ends of the
stress cracks were not obtained as lines. Also, some of
the lines representing stress cracks were not continuous;
some faint cracks were not detectable. This was due to
the image-smoothing step in the algorithm, which tended
to erode the less contrasting portions of the image (the
ends of stress cracks). Therefore, further improvements
in the processing algorithm and/or illumination
conditions are necessary to extract complete stress crack
details.
In all the figures shown here, the kernel edges are
represented by streaks similar to the stress cracks. They
are shown to provide better visualization of the kernel
boundary. However, these edges can be removed by
providing a light seal around the kernel periphery.
CONCLUSIONS
1. White light in the back-lighting mode with a
black-coated background having a 2.4-mm diameter
aperture for the light produced images showing high
contrast between the stress cracks and the rest of the
kernel surface.
2. An image processing algorithm was developed to
produce the effect of high-pass filtering and thus
extracting the pixels representing the stress cracks as
streaks or lines.
3. The algorithm performed very satisfactorily in
detecting stress cracks in 90% of the kernels examined.
4. Double stress cracks were the easiest to detect by
270
1. Berlage, A. G., T. M. Cooper, and R A. Carone. Seed
recognition potential of machine vision systems. ASAE Paper No.
84-3059, ASAE, St. Joseph, MI 49085
2. Gagliardi, G. R., D. SulHvan, and N F. Smith. 1984.
Computer-aided video inspection. Food Technology 42(4): 53-57, 59.
3., Graf, G. L., G. E. Rehkugler, W. F. Millier, and J. A.
Throop. 1981. Automatic detection of surface flaws on apples using
digital image processing. ASAE Paper No. 81-3537, ASAE, St. Jospeh,
MI 49085.
4. Gunasekaran, S. 1985. A laser optical system for
nondestructive detection of corn kernel defects. Unpublished Ph.D.
Thesis, University of Illinois, Urbana.
5. Gunasekaran, S., S. S. Deshpande, M. R. Paulsen, and G. C.
Shove. 1985. Size characterization of stress cracks in corn kernels.
TRANSACTIONS of the ASAE 28(5):1668-1672.
6. Gunasekaran, S. and M. R. Paulsen. 1985. Breakage
resistance of corn as a function of drying rates. TRANSACTIONS of
the ASAE 28(6):2071-2076.
7. Gunasekaran, S. and M. R. Paulsen. 1986. Automatic,
nondestructive detection of corn kernel defects. International advances
in nondestructive testing. Vol 12, Gordon and Breach Science
Publishers, NY. pp 95-116.
8. Gunasekaran, S., M. R. Paulsen, and G. C. Shove. 1986. A
laser optical method for detecting corn kernel defects.
TRANSACTIONS of the ASAE 29(l):294-298, 304.
9. Kahwati, G. and P. Law. 1985. Machine vision system assures
product quality. Research and Development. October, pp 118-119.
10. Matshuisa, T. and A. Hosokawa. 1981. Possibilities of
checking cracks of brown rice using illumination by oblique ray and
image data processing system. J. of the Society of Aguricultual
Machinery 42(4):515-520.
11. Paulsen, M. R. and W. F. McClure. 1985. Illumination for
computer vision systems. ASAE Paper No. 85-3546, ASAE, St. Joseph,
MI 49085.
12. Sarkar, N. and R. R. Wolfe. 1985. Feature extraction
techniques for sorting tomatoes by computer vision. TRANSACTIONS
of the ASAE 28(3):970-974, 979.
13. Shimatachi, Y., Y. Nomura, T. Ide, and O. Itoh. 1982.
Application of pattern measurement technique to fish selection.
Mitsubishi Electronic Technical Report 56(3): 44-48.
14. Sites, P. W. and M. J. Delwiche. 1985. Computer vision to
locate fruit on a tree. ASAE Paper No. 85-3039, ASAE St. Joseph MI
49085.
15. V^hittaker, A. D., G. E. Miles, O. R. Mitchell and L. D.
Gaultney. 1984. Fruit location in partially occluded image. ASAE
Paper No. 84-5511, ASAE, St. Joseph, MI 49085.
APPENDIX
BASIC program to implement the image processing algorithm for
detecting stress cracks in corn kernels in the Intelledex V200 vision system.
10
20
30
40
50
60
70
80
90
100
110
120
130
'Program to process a corn kernel image to detect stress cracks.
'Choose a lower limit for VMAP operation. A lower limit of 20
'is suggested based on kernel image histogram gray-scale levels.
'VMAP enhances contrast between the pixels representing the stress
'cracks and the rest of the pixels by doubling the gray-scale
'value of the pixels defined by a window.
DIM IMAP (63)
INPUT "Enter the lower limit", XL
F O R I = O t o X L : IMAP(I) = 0: NEXTI
FOR I = XL To (XL + 31): IMAP(I) = 2* (I-XL):NEXT I
F O R I = ( X L + 32) To 6 3 : IMAP(I) = 6 3 : NEXT I
VMAP IMAP
TRANSACTIONS of the ASAE
140
150
160
170
180
190
200
210
220
230
240
250
260
270
280
'Store contrast enhanced image in image buffer zero.
'VENHANCE the image three times.
VSIMAGE 0
FOR I = 1 To 3 : VENHANCE 1: Next I
'Subtract gray-scale value of each pixel of the VENHANCEd image
'from corresponding pixels of the image stored in buffer zero
'(VSIMAGE) and add 20. A value of 20 was chosen arbitrarily to
'make the image appear good to the viewer. This, however, wHl
'not affect the end result. You can choose any value.
VSUBTRACT 20, 0
Vol. 30(l):January-February, 1987
290
300
310
320
330
340
350
360
370
380
390
400
410
420
430
'Choose a threshold value of 22 (this value is relative to the
'number 20 used in the VSUBTRACT step). VTHRESH converts all
'pixels of gray-scale value less than the threshold value to pure
'black (0) and the rest of the pixels to pure white (63).
'Compress the VTHRESHed image in bit plane zero.
VTHRESH 22
VCOMPRESS 0
'Restructure the cracks using two directly opposite linear
'structuring elements. Repeat VENHANCE three times.
F O R P 3 To 4: V D I L A T E : N E X T I
F O R I = 1 To 3: VENHANCE 1:NEXT I
END
271