An Image-processing Based Algorithm for Rice Seed

Germination Rate Evaluation

Paween Khoenkaw
Major of Computer Science, Faculty of Science, Maejo University
63, Chiang Mai - Phrao Road, Sansai, Chiang Mai, 50290, THAILAND
[email protected]

Abstract—The seed germination test is an important process classification error, the evaluator must precisely identify the
for successful farming. This process is time-consumed and needed germinated seed and un-germinated seed, this process requires
to be done by an expert. In this paper, the image processing experienced evaluators. Final problem is the evaluator must use
method is applied to measure a germination rate. The proposed a large number of seeds in order to obtain high accuracy result,
method used a digital camera to capture the rice image, the root
this is a time-consuming process.
and seed sections are segmented using the HSV color model. The
seed that has root attached is considered germinated. This method In order to overcome these problems, the image
was evaluated using 40 images of wild rice, the accuracy of our processing based automatic germination rate evaluation was
system of seed number detection is 96.37% and germination rate proposed. Howarth and Stanwood [1] was proposed the
evaluation is 98.41%. germination rate evaluation system for sorghum and Lettuce,
this system provided lighting and photographing system,
Keywords image processing; germination test; rice seed; pattern recognition algorithm is used for root detection,
however this system only recognize root that is grown in a
I. INTRODUCTION
straight line. Y.Sako[2] also proposed a low cost germination
The seed selection is the important part of the rate evaluation system for Lettuce, the best result of this system
agriculture, the quality of seed is one of the most important is nearly identical to the result done by human, however the
factors in successful farming. Good seed is the key to the average error is high as 14.7%. R Ureña [3] used a color CCD
successful farming, so the seed quality measurement is the camera to capture the color image and use the color to detect
pillar of the agriculture. The survival rate and seed storage the cotyledons, the fuzzy logic also used in the research to
quality can be predicted by seed quality measurement, it can be decide that the seed is germinated or not. The disadvantage of
used to plan, predict the crop that is profitable to harvest or not. this research is it has to wait for the leaf to sprout to get the
One of the seed quality measurement is the result. A.L. Hoffmaster[4] developed the germinated rate
“Germination test”, this method plays an important role of the evaluation system for soybean, this system capture image of
farmer who crops depend on good seed such as rice, corn, soybean that germinated on the white tissue paper. This system
sorghum and beans. Because the germination rate is the best evaluated the germinated seed by using statistical methods to
indicator for good quality seed. The seed producers also use analyze the shape and length of the soybean.
germination rate to grade the product, farmers also use The challenge of image processing based germinated
germination rate to determine how many seed is needed per rate evaluation of rice is contained 3 challenges. First, the rice
area. Because high germination rate seed can be reduced the seed is small and difficult to photograph in fine detail. The
amount of seed needed per area, so the farming cost can be second challenge is the root of sprout seed is not a straight line
significantly reduced. and likely to tangle with other seed, this makes the root difficult
Rice is the most important crop in Thailand, many to detect and segment. The final challenge is the rice germinated
families depended on this crop as the main source of income. and grow roots in difference rate, in one sample the root of each
The seed quality evaluation is a very important step to grow this seed can be varied in length up to 4 times. The example of
crop, because the quality of this seed is varied based on seed germinated rice seeds in one sample is shown in Fig. 1.
storage condition. In this research, we proposed the novel image processing based
Traditionally, the germination rate is determined by automatic rice seed germination rate evaluation system that can
putting 100 seeds on the damp paper towel, after 7 days, then overcome these three challenges. The propose model is
the germinated seeds are counted and this value is the described in section 2, the evaluation results are described in
germination rate of that sample. In order to get higher accurate section 3, and the conclusion is described in section 4.
value, this step need to be repeated with the larger number of
seeds.
There are 3 main problems with traditional
germination rate evaluation method. The first problem is the
counting, the evaluator must count the seed twice (before
germinated and after germinated) if human error happening in
this process the error will be double. The second problem is the

978-1-5090-4420-7/16/$31.00 ©2016 IEEE

978-1-5090-4420-7/16/$31.00 ©2016 IEEE

 Digital Camera

1.5 Meters

Light Light

Rice Seeds
Fig. 1 Variety germination stages of seeds in one sample
Fig. 3 Camera and light configuration for seeds photographing
II. PROPOSED METHOD
The proposed image processing part is composed of
The principle of our proposed system is to classify three main modules; background subtraction module, seed
seed as germinated or not germinated by looking for the root or segmentation module and root segmentation module. The detail
coleoptile and seed, if the seed has root attach, then seed is of the proposed algorithm is illustrated in Fig. 4.
germinated and vice versa. The component of germinated rice The first part of the image processing process is the
seed is illustrated in Fig. 2 conversion the RGB image captured by a digital camera into
HSV color space. Then, the background subtraction is applied
Primary leaf
by eliminating dark pixels using equation (1), where 𝐼𝑚𝐻𝑆𝑉 is
an input image in HSV format, 𝐼𝑚𝑉 is a V channel of 𝐼𝑚𝐻𝑆𝑉 ,
Coleoptile 𝜃𝑓 is the minimum brightness of foreground image and 𝐹𝑔𝐻𝑆𝑉
is the segmented foreground image in HSV color space.

Seed 𝐼𝑚 (𝑥, 𝑦) ; 𝑤ℎ𝑒𝑟𝑒 𝐼𝑚𝑉 (𝑥, 𝑦) > 𝜃𝑓

𝐹𝑔𝐻𝑆𝑉 (𝑥, 𝑦) = { 𝐻𝑆𝑉
0 ; 𝑜𝑡ℎ𝑒𝑟𝑠

Primary root
Next step is the root and coleoptile segmentation, the
purpose of this step is to segment only the root and coleoptile
image from the foreground that contained both seeds and roots.
Fig. 2 Component of germinated rice seed To extract only root and coleoptile the Hue and Saturate are
used to locate the pixels that have white color, because white is
There are two important parts in this proposed method,
the color of plant root and coleoptile. The root and coleoptile
rice seed photographing part and image processing part. First,
segmentation process are described in equation (2), where 𝐹𝑔𝑆
the rice seeds are dried with cold air and spread evenly and
photograph in the dark tray. The tray is painted with the dark is a saturate channel of foreground image, 𝜃𝑟1 and 𝜃𝑟2 are
color. The camera and light source configuration are shown in threshold values and 𝑅𝑜𝑜𝑡 is the binary image contained only
Fig. 3, high resolution DSLR is used to photograph the rice root part.
seeds. Two light sources are used to illuminate the rice seed, the
lights are pointed with shallow angle to prevent the light
reflection on the seed containing tray. However, this light 1 ; 𝑤ℎ𝑒𝑟𝑒 𝜃𝑟1 < 𝐹𝑔𝑆 (𝑥, 𝑦) < 𝜃𝑟2
𝑅𝑜𝑜𝑡(𝑥, 𝑦) = {
configuration creates the shadows of the rice seeds. But these 0 ; 𝑜𝑡ℎ𝑒𝑟𝑠
shadows can remove later in the process of image processing.
A long focal-length photographic lens is used to avoid image
distortion, because rice seed is small, then the image distortion Next step is the seed extraction, this step produced the
must be avoided. image that only contained seed by subtract foreground image
with the root image as describe in equation (3), where
⌈𝐹𝑔𝑉 (𝑥, 𝑦)⌉ is the ceiling function applied to value channel of
foreground image to convert grayscale to binary image and foreground model generated using equation (1), and the whole
𝑆𝑒𝑒𝑑(𝑥, 𝑦) is the binary image contained only seed and seed and root segmented result is shown in Fig. 5F. Fig. 5G is
coleoptile. the segmented root image using equation (2) and Fig. 5H is the
segmented seed image using equation (3).

𝑆𝑒𝑒𝑑(𝑥, 𝑦) = ⌈𝐹𝑔𝑉 (𝑥, 𝑦)⌉ − 𝑅𝑜𝑜𝑡 Color Image

Next step is the germinate detection, it is done by

looking for the root or coleoptile which has a connection to the Convert RGB
to HSV
seed. The process is started by partition rice seed image into
blobs that contain only individual seed per blob using an Image
Labelling algorithm. However, if seeds are put too close
together they are likely to be labelled as the single grain. This
problem is solved using seed number estimation to estimate the
Background
seed count in each blob. Seed number estimation is done by Subtraction
comparing the sum of the pixel value with the number of
reference single grain as described in equation (4), where 𝐴 is
a sum of pixel value in the blob and 𝑆𝑒𝑒𝑑𝑟𝑒𝑓 is a reference sum
pixel value of a single seed image.
Root Seed
Segmentation Segmentation
𝐴
𝑆𝑒𝑒𝑑𝐶𝑜𝑢𝑛𝑡 = ⌈ ⌉
𝑆𝑒𝑒𝑑𝑟𝑒𝑓

Then, the root searching windows are created using Seed

Labeling
position of each blob plus 𝑟 pixels on every side. The
germination is detected by put the root searching window in the
root image (𝑅𝑜𝑜𝑡), if sum of value in the window is greater
than 𝜃𝑔𝑒𝑟𝑚𝑖𝑛𝑎𝑡𝑒𝑑 then consider root is founded.
If the root is found in this window, then the seed in this Seed number
blob is considered germinate. However the blob that contained estimation

more than one seed is not used because it cannot determine

which seed is germinate, this issue will be discussed in the detail
in the experiments section.
Finally, the germination rated is calculated by Yes
Is root
equation (5), where 𝑛 is germinate seed count and 𝑜 is non- connected ?
germinate seed count.
No
No Yes
Is number of
𝑜 seed >1 ?
𝐺𝑒𝑟𝑚𝑖𝑛𝑎𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒 = × 100
𝑛+𝑜

n=number of seed o=number of seed p=number of seed

𝑝 in blob in blob in blob
′
𝐺𝑒𝑟𝑚𝑖𝑛𝑎𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒 = × 100
𝑛+𝑝

𝑜+𝑝
𝐺𝑒𝑟𝑚𝑖𝑛𝑎𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒 ′′ = × 100
𝑛+𝑜+𝑝 Germination
Rate=n/(o+n)x100

The example output of each step in this proposed Fig. 4 The proposed seed germination rate evaluation algorithm
algorithm is shown in Fig. 5 , where Fig. 5A is the original input
image, Fig. 5B, Fig. 5C, Fig. 5D is the component of Hue,
Saturation, Value from color separation module. Fig. 5E is the
 Fig. 5J is shown the classification result where The image used in this experiment is resized to
detected seed marked with blue rectangle. The green and red 1020x1212 pixels. The preliminary experiments are conducted
rectangle are the search window used to find the root. These to determine the best threshold values and the results are 𝜃𝑓 =
windows are painted in green if the root is founded and painted 0.3, 𝜃𝑟1 = 0, 𝜃𝑟2 = 0.3, 𝑆𝑒𝑒𝑑𝑟𝑒𝑓 = 1274, 𝜃𝑔𝑒𝑟𝑚𝑖𝑛𝑎𝑡𝑒𝑑 =90.
in red if the root is not found. These thresholds are used for performance evaluation of this
algorithm.
The 40 images are used to evaluate the accuracy of
number of seed estimation and germination rate. The four
parameters of this algorithm were evaluated. First parameter is
seed number estimation. It is the ability to locate the rice seed
include the germinated seed and non-germinate seed. Method I,
A B equation (5) is the ability to detect the germination rate by only
considered blob that contained a single seed. Method II,
equation (6) is the ability to detect germination rate by only
considered blob that contained multiple seeds. In this case, the
algorithm is considered all seed in in the blob are germinated if
the root is found, or consider all seed are non-germinated if root
is not found. Method III, equation (7) is the ability to detect
C D germinated rate by combining Method I and Method II. The
results of this algorithm are compared with ground truth
generated by the expert.

𝑆𝑒𝑒𝑑𝑁𝑢𝑚𝑏𝑒𝑟𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑖𝑜𝑛 = 𝑛 + 𝑜 + 𝑝
E F
The experimental results are shown in Table 1, it is
shown that the best solution is Method I. The best accuracy of
germination rate evaluation is 98.41% ±1.87.

G H
TABLE 1 EVALUATION RESULTS
Parameter Accuracy SD
Seed number estimation 96.37% 3.94
Method I 98.41% 1.87
Method II 91.65% 11.10
I J Method III 95.69% 7.80
Fig. 5 The sample image output of each step in the proposed algorithm
The important factor that causing Method I to gain the
III. EXPERIMENTS AND RESULTS highest accuracy is the seed location. The seeds that located far
or isolated from other seed are not likely to give error in
The rice seed used in these experiments is the Oryza segmentation or detection process. Since the rice seeds are
rufipogon (wild rice) which is widely used for animal feed, uniformly spread while the photograph is taken, then this type
seeds are 2 years old stock which are purchased from a local of seed can be used to represent the whole population of rice in
store. This is ensured the diversity mixed of germinate and non- that tray.
germinate seed. The seeds are put on the damp towel paper for
8 days. Then spread evenly on a dark ceramic tray and
photograph. The flat black WIN acrylic lacquer (color code:
202 400CC) was used to paint the photographing tray, this paint
is dark black and low light reflection.
The Canon EOS400D with telephoto lens 55-120mm
is used in this experiment. The image was taken with ISO 100;
shutter speed 1/125 and aperture value of 5.0. The white balance
correction was done automatically by the camera using a grey
card as a reference.
 extracted using the HSV color space. This algorithm counts the
seeds and classify them as germinate or non-germinate. The
seed that has root attached is considered germinated. This
algorithm was evaluated using 40 rice image. The Oryza
rufipogon was used in these experiments. The results are
compared with ground truth generated by the expert. The
accuracy of our system in seed counting is 96.37% ± 3.94 and
germination rate evaluation is 98.41% ±1.87. This algorithm
B helps farmer to reduce time consuming process of counting and
A sorting the rice seeds in the germination rate evaluation process.
Fig. 6 Error cause by tangled root ACKNOWLEDGMENT
In this experiment, we found that the accuracy of this This research was partially supported by The Office of
algorithm has depended on the condition of light and seed Agricultural Research and Extension Maejo University, and
moisture. Lighting is effected the root and seed segmentation Unit of Excellence for Intelligent Digital Innovation Maejo
algorithm, intense light can cause the color clipping, that result University (INTNIN Laboratory).
in the camera cannot capture the yellow color. Moisture is
effected seed’s light reflex, wet seed or water droplet on the
REFERENCES
seed can cause the light to reflect to the camera and cause color
clipping. The segmentation algorithm can mistake the water 1 Howarth, M.S., and Stanwood, P.C.: ‘Measurement of
droplet as root. The focusing mechanism of the lens used in seedling growth rate by machine vision’, in Editor
this experiment is causing the focal length to shift from the (Ed.)^(Eds.): ‘Book Measurement of seedling growth
desired position. That resulting in the size of the seed is larger rate by machine vision’ (1993, edn.), pp. 185-194
or smaller than usual, using a fixed focus lens is highly 2 Y. Sako, M.B.M., K. Fujimura, A. F. Evans, and M.
recommended. A. Bennett: ‘A System for Automated Seed Vigor
Now we are currently developing the automatic light Assessment’, 2001
adjustment hardware and algorithm to overcome the lighting 3 R Ureña, F.R.g., M Berenguel: ‘A machine vision
problem and the preliminary are promising. system for seeds quality evaluation using fuzzy logic’,
Computers and Electronics in Agriculture 2001, 32,
pp. 1-20
IV. CONCLUSION 4 A. L. Hoffmaster , K.F., M. B. Mcdonald, M. A.
In this paper, we proposed a new algorithm for image Bennett: ‘An Automated System For Vigor Testing
processing based rice seed germination rate evaluation. The Three-Day-Old Soybean Seedlings ’, 2003
images of the test rice seeds were taken by a digital camera on
a dark color tray. Then the seeds, roots and coleoptile are