IMAGING TECHNIQUES TO ENHANCE

THE PRESERVATION AND UTILIZATION

OF SEED GERMPLASM
M. Scott Howarth and Phillip C. Stanwoodl

ABSTRACT
Several image processing techniques are belng developed at the
National Seed Storage Laboratory to enhance the preservation and
utilization of seed germplasm.Two projects investigate the potential
of this technology in controlling and providing additional informa-
tion in germination testing, seedung growth rate analysis and tetra-
zolium testing. Another thrust of the research has been to develop
an image-based database. Investigations in this study concentrate
on the development of features which describe physical character-
istics of seed, An object-oriented database was developed within
the Windows 3.1 operating environment. This database Is unique
because it is capable of displaying many types of information, tor
example seed images, graphs, or color pads. Machine vision and
image analysis have shown promise and many new applications of
this technology are being developed.

Additional index words: machine vision, seed vigor, tetrazolium test-

ing seed characteristics, slant board.

INTRODUCTION

he United States National Plant Germplasm System (NPGS), which

T
for
includes the National Seed Storage Laboratory (NSSL), is responsible
long-term preservation. collection, propagation and distribution of
the
pCani genetic resources (Shands, 1990). NPGS collections contain major
agricultural crops such as corn, wheat, rice, and other plant collections such
as vegetables and ornamentals. These collections contain wild species as
well as developed cultivars. Most of this germplasm is held as seeds.
The record keeping for germplasm preservation is a never ending task.
Agronomists, plant pathologists, plant breeders, conservationists and
many other scientists are faced wlth the challenge of creating, updating,
and maintaining germplasm databases. The Germplasm Resources Infar-
mation Network (GRIN) database was developed in order to utilize and
maintain germplasm. The task of maintaining this database is the respon-
sibility of the National Germplasm Resources Laboratory of the United
States Department of Agriculture, Beltsville, MD, Some of the information
within GRIN describes physical characteristics of the seeds. These char-
acteristics include hundred seed weight, shape, color, and texture. With
the exception of hundred seed weight. all of these characteristics involve
subjective evaluations. In many accessions genetic variability is repre-

1 Agricultural Engineer and Research Agronomist, USDA-ARS, National Seed Storage

Laboratory, 1111 South Mason, Fort Collins, CO 80521-4500.
 JOURNAL OF SEED TECHNOLOGY 55

sented by differences in seed size (length, width, and thickness), color,

texture, etc.These differences are not recorded in GRIN. Much of this vari-
ability could be expressed by color images and other extracted features.
Seed deterioration is another problem facing the NPGS and NSSL. This
has plagued mankind since the cultivation of plants began and has been
a major concern for the NSSL. The main objective at NSSL is the
long-term preservation of plant germplasm for the United States. Plant
germplasm is stored at sub-zero temperatures, and the questions often
asked are "How long can a seed sample be stored under these conditions
and maintain viability?" and "Why do these seeds deteriorate?" Obviously,
it is important to measure the viability of the seeds both before and during
storage. Before seeds are placed into storage, the initial viability is deter-
mined. During storage, seeds samples are tested to determine the rate of
deterioration and if regeneration is needed.
Seed vigor and germination tests have traditionally been used to deter-
mine deterioration of seed samples. A seed vigor test describes the seed
potential to emerge rapidly and produce a mature crop under certain field
conditions (Heydecker, 1972). The test must be performed in an objective
manner to maintain a common language between all testing entities.
Although AOSA (1988) attempts to provide guidelines for testing seeds, it
also requires subjective interpretation of the results. This can cause dis-
crepancy among experienced analysts.
The current seed viability tests are labor intensive and subjective. The
standard germination test (AOSA, 1988) determines the percent germi-
nation of a seed sample, and takes 4 to 28 days or longer to complete. If
the dormancy of the seeds is not overcome, percent germination will not
reflect the true viability. Tetrazolium (TZ) testing can be used to detect
viability of dormant seeds. Results from TZ tests often can be obtained
within 24 hours. However, due to the high level of subjectivity in assessing
color development, the TZ test cannot replace the standard germination
test. Machine vision is a tool that might assist researchers and seed
analysts to measure and predict seed viability.
There have been numerous studies on the use of machine vision for
improving agriculture product quality and most has concentrated on color
and blemish recognition. Several researchers have worked on detecting
bruises on apples (Taylor et al., 1984; Taylor and Rehkuglar, 1985; and
Brown et al., 1974). Computer Recognition System developed a orange
grading system which sorted oranges into three classes (Harris, 1988).
Howarth (1991) developed a machine vision system to grade fresh market
carrots. This system identified blemishes caused by rot but also graded
the carrots based on several shape parameters.
Many image processing applications have been developed to measure
seed characteristics. Churchill et. al. (1990) developed a machine vision
system to measure seed dimensions of tall fescue, orchardgrass and
perennial ryegrass. Texture and color image processing techniques were
developed to identify weed seeds (Petersen and Krutz, 1992). Feature
extraction techniques using fractals were used to describe the shape of
ear corn (Panigrahi and Misra, 1990). Image analysis has been identified
as a tool that could be used for taxonomy (Draper and Travis, 1984). Most
56 VOLUME 17, NUMBER 2,1993
recent attempts to measure seed shape by machine vision have primarily
concentrated on specific varieties or on classifying the difference between
a small group of varieties.
McCormac st al. (1990) investigatedimage analysis for measuringthe root
length of lettuce using a slant board test. Mwvwer, several problems were
identified with this technique, First, root length was only measured after the
slant board test was completed. Secondly, only linear length was measured;
therefore, the true length of the root may not have been measured.
The objectives of tire computer imaging research at the NSSL are to:
1) increase the efficiency and accuracy of evaluating stored seed germ-
plasm, 2) reduce the amount of personnel resources needed to monitor
seed viability, and 3) provide information about seed accessions that is not
currently available and to make this information readily available to users
of germplasm.

IMAGE PROCESSING SYSTEM2

The basic components of a computer imaging system consist of a
imaging sensor, an image frame grabber, a micro-computer, and a video
monitor. The imaging sensor converts electromagnetic energy into an
analog signal. The most common sensor is a CCD (Charged Coupted
Device) carnera.The image frame grabber converts the analog signal from
the imaging sensor into a digital signal. The digital signal is scored in a
buffer on the frame grabber and can be accessed via the computer data
bus. The digital signals are organized into a matrix of pixels. The value at
each pixel is relative to the reflected light at a specific location within the
sensor's field of view. This is generally a value between 0 and 255; 0
corresponds to no light reflected or black and 255 corresponds with a
maximum reflectance or white.
The computer host systems in the NSSL imaging laboratory are PC-based
386 and 486 systems.The two image frame grabbers utijized are a Matrox
640 PIP and a Matrox MVP-AT.The Matrox 640 PIP image processing board
processes and stores gray scale images in a 640 x 480 array. The black and
white camera connected to this board is a Cohu camera with a 50 mm lens.
Also included with this camera setup is a microscope adapter. The Matrox
MVP-AT is a multiuse frame grabber. This board can be configured as either
a color or gray scale imaging board. In the color mode, the array size of the
board is 512 x 480 and in the gray scale mode, the array size can be set in
several different sizes: 4 - 512 x 480 to 1 - 1024 x 1024.Two color cameras
can be utilized with this board, a Cohu color camera and a Sony 3 CCD color
video camera. The color Cohu camera, like the gray scale camera, accepts
'C' mount lenses. A 50mm lens was used. The Sony camera used a bayonet
mount, and a telephoto lens (7.5 to 90 mm) was used. A microscope adap-
tor was available for use with an Olympus microscope.
The two operating environments were DOS 5.0 and Windows 3.1. Most
of the developmental work was done in the Windows 3.1 environment and
once a concept was developed, it was ported into the DOS 5.0 environ-

Mention of company or trade name does not imply endorsement of the products by the
USDA-ARS. It is for purpose of description only.
 JOURNAL OF SEED TECHNOLOGY 57
Stults Germination
Chamber

Solid State Relay

-I
Figure 1. Schematic of machine vision system for root length measurement.

ment in which many of the imaging operations were performed on the

imaging boards and at fast rates. The developmental package used within
the Windows 3.1 environment was OPTIMAS 3.1. This package was use-
ful for applying different filtering and other image processing techniques.
The OPTIMAS software also included an Analytical Language for Images
(ALI) which was used to write macros. Turbo C++ by Borland was used in
the Windows 3.1 environment to develop Windows application for color
image processing. In the DOS 5.0 environment, Microsoft C version 5.0
was used to implement algorithms used to make image measurements.

IMAGING PROJECTS
The four major projects at the NSSL are: 1) to develop a machine vision
system to measure seedling growth rate and determine if this information
can be used to predict seed viability, 2) to develop a machine vision to
assist the seed analyst to perform a non-subjective tetrazolium viability
test, 3) to develop image processing and feature extraction techniques to
measure the physical characteristics of seeds, and 4) to develop an
image-based database for plant germplasm. An outline of the concepts for
these projects and the latest advances follow.

Seed Growth Rate Test

The objectives of this study were 1) to develop feature extraction tech-
niques needed to measure seedling growth rate, 2) to develop a machine
vision system capable of monitoring seedling growth rate, and 3) to deter-
mine the reliability of the machine vision system. A machine vision system
was developed to measure root growth rate over the entire germination
period. The system schematic is shown in Figure 1. A Cohu black and
 VOLUME 17, NUMBER 2,1993

Flgure 2. A typical Image obtained durlng testlng of ten sorght~mseeds. Note the two
pins seen as round white objects.The sorghum seeds were planted in a row between
these pins. Roots are shown attached to the blotter paper extending down from the
pin location and the stem portlon extends up from the pin location.

white camera was set outside a Stults germinatlon chamber. This cham-
ber provided the necessary environmentat conditions to run a standard
germination test. The chamber remained dark throughout the test except
when images were obtained during which tights remained en for approxi-
mately 30 seconds every hour. The images were stored on the hard disk
and analyzed later. An example of an image is shown in Figure 2.
The steps to lacate and measure the root growth included calibration,
location, and measurement. Two calibration measurements were made in
each image, positional and size. As shown in Figure 2, two circutar posi-
tioning pins were used to perform the calibration rneasurement.These cir-
cular pins were located, and the position of the pins and distance between
the two pins were used for calibration. Once the pins were located, roots
were located and the length was measured. This was completed for all
images and the growth raZe curves for lettuce and sorghum are shown in
Figure 3. These measurements compared similar to human measure-
ments. The average error between the two measurement systems were
-0,13cm for the lettuce test and -0.07 cm for the sorghum test. For more
detail on the procedures and results, see Howarth and Stanwood (1992a).
This system analyzes ten seeds every germination period. An auto-
mated system has been developed and is under construction which will
present 50 slant boards to the camera. This will allow 500 seeds to be
analyzed every germination period.This wiH make this system attractive to
the NSSL research and preservation units and to the seed industry.
 JOURNAL OF SEED TECHNOLOGY

- Root 1
Root 2
. - - . Rmtot3
....... Root 4
--- nwt5
.- - . Root 6
Root7
----- nmta
- noot 10

Time, hours

Time, hours
Figure 3. Illustration of ten root growth rate curves measured by the machine vision
system: A) lettuce and B) sorghum.

Tetrazolium Seed Viability Test

In order to automate the TZ viability test, it was necessary to develop
color feature extraction and classification techniques to automate the clas-
sification of TZ stained corn seeds. A machine vision system was devel-
oped using the color image processing system and is fully described in
Howarth and Stanwood (1992b). The ratio between stained area and total
corn area was used as the primary feature. The stained area was located
and measured by subtracting the red image plane from the green plane.
60 VOLUME 17, NUMBER 2,1993

Normal corn samples

Dead corn samples

000 005 0.10 0.15 020 0.25 0.30 035 0.40 945 0.54 0.55
TZ staining ratlo

Figure 4.TZ staining ratio histogram showing the distribution of normal and abnormal
or dead corn kernels.

This produced an enhanced image which was thresholded. Thresholding

is a procedure in which an optimum intensity value (also called the thresh-
old) is selected. Then, the intensity value of all pixels within the image are
compared to the threshold value. All intensity values above the threshold
are classified object and value below the threshold value are classified
background. This transformed image is now considered a binary image.
Then, this binary image was organized using the connected components
algorithm. The total corn area was segmented from the background by
thresholding the intensity of the image. The intensity is the average of the
red, green and blue (RGB) image planes. The classification histogram was
constructed from the training data and is shown in Figure 4. This his-
togram shows the distribution of normal and abnormal or dead corn ker-
nels. Based on this information, a Bayes classifier was developed. Overall,
the misclassification rate for the training tests was 8 of 80 samples
(lO.OOh).For the validation set, 7 samples were mis-classified out of 67
(10.4%). However, all these errors were Type II errors. The Type II error in
this case is the dead or abnormal corn kernels that were incorrectly clas-
sified normal. This was caused by the machine vision not being robust
enough to identify special cases. These cases were when a specific and
important region within the stained embryo was not stained. By automati-
cally identifying and locating these non-stained regions, the current sys-
tem could reach its full potential.

Physical Characterization of Seeds

Much of the seed characteristic information within the GRIN database is
based on subjective evaluation. A large degree of variability can be repre-
 JOURNAL OF SEED TECHNOLOGY 61

Figure 5. Sample image obtained from the 3-D imaging system.The center image rep-
resents the top view of the seed. The top and bottom images are side views of the
seed.The right and left images are the end views of the seed.

sented within a seed sample. Therefore, a tool was needed to evaluate

physical characteristics of seed in a non-subjective procedure. As men-
tioned earlier, many image processing applications have been developed
to measure seed characteristics; however, all have concentrated on two
dimensional images. In many cases, some characteristics can be obscured
or hidden from view using only two dimensional images.
In order to collect three dimensional information, a machine vision
system was developed to measure this information with a two dimension-
al imaging system. A mirror mount was constructed and this mount was
fabricated from aluminum so that four mirrors could be position in pyra-
midal fashion about a seed positioning stand. The mirrors are positioned
at 45" incident to the imaging plane. An example of an image obtained
from this system is shown in Figure 5. As shown in Figure 5, the center
corresponds to the top of the seed and about the center are four reflected
images of the two side and two end views of the seed.
Three color pads also shown in Figure 5 were used to locate area of
interest (AOI) regions and to spacially calibrate the images. These pads
were located using a circle detection algorithm (Howarth and Stanwood,
1992~).Once these color pads were located, five A01 regions were
located and several size, shape and color features were calculated from
each region. The size features were length, width, and height. The shape
features were form factor, circularity, and centroidal profile. The color fea-
tures were the RGB statistics (i.e. average and standard deviation) and
62 VOLUME 17, NUMBER 2,1993

RGB histograms. For more specific information regarding this work, refer
to Howarth and Stanwood (1992C).

Image-Based Plant Germplasm Database

Traditional approach to database programming and management has
revolved around the textual data with which the program and rnanage-
msnt environment must deal. Recent interest has focused on non-textual
information (i.e. images, graphs, etc.). Virtually all proposals from the
database community for management of non-textual information use
object-oriented techniques (Grosky and Mehrotra, 1989). To be able to
handle images, graphics, and other non-textual information, it appears
that an object-oriented database environment is the most efficient and
attractive technique.
Object-oriented programming is a paradigm which is distinguished by its
representation of the relationship between data and programs (Engel et al.
1989).Traditional programming languages like FORTRAN and C operate in
a procedural environment. In this proc~duralenvironment, there is typically
a main function and additional functions that are called from the main func-
tion (Pappas and Murray, 1992). This is commonly referred to as the
top-down approach,An object-oriented program is constructed from a group
of related "objects". These objects form a hierarchy which makes an
object-oriented program more modular; therefore, new objects and methods
can be added with relativelyfew unanticipatedside effects (Fernhout, 1989).
Where traditional databases require all data elements to have similar for-
mats, an object oriented database can tailor data structures to describe the
information being stored. This allows the database to contain a variety of
information, from text to graphics to computer simulations (Beck, 1989).
Therefore, this type of database has the ability to use an abject-oriented
programming language to describe different data structures by defining
classes that are individuaZly tailored to the data (Freeman and Field, '1992).
An object-oriented color image database of plant germplasm was devel-
oped in the Windows 3.1 operating environment. The images were dis-
played as full color (over 16 million colors) bitmaps using a super VGA
card and monitor. The information within this database included the
images, passport and seed characteristics downloaded from GRIN, and
physical characteristic seed data collected from the 3-D imaging system.
Several types of objects were created to handle this variety of data.These
included objects to handla and display bitmaps, graphical and textual
information, and to generate and display color pads and representative
shape figures. Presently, 33 accessions of Phaseolus, 20 accessions of
endangered species seed, and several other samples of various seeds
have been input into this database. This database may provide curators,
researchers and plant breeders with a useful tool for visually inspecting
current seed of plant germplasm.

FUTURE DIRECTION
During the last few years, the projects discussed have provided the
NSSL with a sound foundation for future development. Several areas must
be investigated so that the current projects may reach full potential.
 JOURNAL OF SEED TECHNOLOGY 63

Discussions with other researchers at the NSSL and Colorado State

University have begun to expand the scope of the imaging research.
For current and ongoing projects, several major concerns that will be
addressed are as follows. The first projects slated for the next year is
the testing of the SGRobot, which is the automated system for mea-
suring seedling growth rate. Furthermore, tests are being designed to
understand the biological meaning of different root growth rates and to
determine the link between seed growth rate, seed vigor, and seed
longevity. The TZ project will be improved to distinguish special cases of
staining.This, along with testing of many cultivars of corn, will enhance the
current system to provide seed analysts with a tool to quickly measure
seed viability. For the seed characterization project, it is important to
develop a system which is more robust and more tolerant of positional
bias. This would alleviate positional error and simplify image acquisition.
Other important aspects of this system that will be investigated are
improvements in the lighting and focusing systems to help to reduce over-
all system error. To realize the full potential of the image-based database,
several problems will need to be addressed. First, a technique to com-
press the images must be incorporated into this database, and, secondly,
the method of distribution must be addressed. As these operations are
installed, plant breeders, seed scientists, and other researchers can
began to use this system and comment on its abilities. Their comments
will be useful for development of future versions.
One of the future imaging research projects slated is pollen germination
counts. The current procedure is very time consuming and causes eye
fatigue. An imaging system could help to count germinated and non-
germinated pollen. As well, a technique to observe and measure germi-
nating pollen tubes non-invasively will also be investigated.
The imaging technology shows enormous potential for the enhance-
ment and development of non-invasive biosensors. These sensor can be
used as tools to perform tedious and subjective tasks and to help unlock
the mystery of seedling vigor.

REFERENCES
1. Association of Official Seed Analysts (AOSA). 1988. Rules for testing
seeds. Journal of Seed Technology, 12(3).
2. Beck, H.W. 1989. Database management. In: Knowledge Engineering
in Agriculture, J.R. Barrett and D.D. Jones (eds.), American Society of
Agricultural Engineers, St. Joseph, MI 49085-9659. pp 83-1 15.
3. Bowman, H. 1990.The use of hydrogen peroxide as a bleaching agent
in the TZ testing of smooth bromegrass seed. Newsletter of AOSA
44(1):26-29.
4. Brown, G.K., L.J. Segerlind, and R. Summitt. 1974. Near-infrared
reflectance of bruised apples. Transactions of the ASAE. 17(1):17- 19.
5. Churchill, D.B., D.M. Bilsland and T.M. Cooper. 1990. Feature extrac-
tion techniques for corn germplasm by color computer vision. ASAE
paper No. 90-7519.
6. Draper, S.R. and A.J. Travis, 1984. Preliminary observations with a
computer based system for analysis of the shape of seeds and
64 VOLUME 17, NUMBER 2,1993

vegetative structures. Journal of the National Institute of Agricultural

Botany 16:387-395.
7. Engel, B.A., R.H. Thieme, and A.D. Whittaker. 1989. Representation
and reasoning. In: Knowledge Engineering in Agriculture, J.R. Barrett
and D.D. Jones (eds.), American Society of Agricultural Engineers, St.
Joseph, MI 49085-9659. pp 47-76.
8. Fernhout, PD. 1989. Simulating interacting intelligent objects in C. Al
Expert 4(1):38-46.
9. Freeman, S.A. and W.E. Field. 1992. Object-oriented databases for
rural assistive technology. American Society of Agricultural Engineers
Paper 92-3504. St. Joseph, MI 49085-9659.
10. Grosky, W.I. and R. Mehrotra. 1989. Image database management.
Computer 22(12):7-8.
11. Harris, M.L. 1988. Machine vision for grading of oranges. Advanced
Imaging. Nov./Dec. No. 15:40-43.
12. Heydecker, W. 1972. Vigour. In: Viability of Seeds. E.H. Roberts (ed).
Syracuse University Press. pp. 209-252.
13. Howarth, M.S. 1991. Machine vision inspection of fresh market car-
rots. Ph.D. Dissertation, Texas A&M University.
14. Howarth, M.S. and P.C. Stanwood. 1993a. Measurement of seedling
growth rate by machine vision. Transactions of the ASAE 36(3):959-
963.
15. Howarth, M.S. and P.C. Stanwood. 1993b.Tetrazolium staining viability
seed test using color image processing. Transactions of the ASAE
36(6):1937- 1940.
16. Howarth, M.S.and P.C. Stanwood. 1 9 9 2 ~3-D . color imaging of seed
using 2-I3 images. ASAE paper No. 923590, St. Joseph, MI.
17. McCorrnac, A.C,, P.D. Keefe, and S.R. Draper. 1990. Automated vigour
testing of field vegetables using image analysis. Seed Science and
Technology 18: 103-1 12.
18. Panigrahi, S. and M.K. Misra. 1990. Feature extraction techniques for
corn germplasm by color computer vision. ASAE paper No.907050.
St. Joseph, MI.
19. Pappas, C.H. and W.H. Murray. 1992. 'Barland C++ Handbook, Second
Edition, McGraw-Hill, Inc, Berkeley, CA 94710.
20. Petersen, P.E.H. and G.W. Krutz. 1992. Automatic identification of
weed seeds by colour machine vision. Seed Science and Technology
20: 193-208.
21. Shands, H.L. 1990. Plant genetic resources conservation: The role of
the gene bank in delivering useful genetic materials to the research
scientist. Journal of Heredity 81:7- 10.
22. Taylor, R.E., G.E. Rehkugler, and J.A. Throop. 1984. Apple bruise
detection using a digital line scan camera system. In: Agricultural
Electronics - 1983 and Beyond. American Society of Agricultural
Engineers, St. Joseph, MI. pp. 652-662.
23. Taylor, R.E. and G.E. Rehkugler. 1985. Development of a system for
automated detection of apple bruises. In: Agri-Mation 1. American
Society of Agricultural Engineers, St. Joseph, MI. pp. 53-62.