Non-destructive Measurement of Vegetable Seedling Leaf Area

Non-destructive Measurement of Vegetable Seedling
Leaf Area using Elliptical Hough Transform
Chung-Fang Chien, Ta-Te Lin
Department of Bio-Industrial Mechatronics Engineering,
National Taiwan University,
Taipei, Taiwan, ROC
INTRODUCTION
• Traditionally
– measuring dry weight, fresh weight, plant
height and coverage to represent plant
growing stages
– destructive and laborious
OBJECTIVES
•
•
•
•
•
Non-destructive
Fast and easy
Image processing
To count the seedling leaf number
To measure and estimate the leaf dimension
(area and perimeter)
MATERIALS AND METHODS
• Materials
Cabbage, Chinese cabbage, Broccoli
Growing at 25℃(day) / 20℃(night)
10 to 30 days after seeding
1 to 4 leaves
MATERIALS AND METHODS
• Methods
Hough transform for ellipses
Focusing
Morphological transformation
Hough transform for ellipses
r 
a 2b2
b 2 cos 2θ  a 2sin 2θ
• 5-dimensional parameter space
• Only the object pixels
• Vote for thresholding
Focusing
Lower resolutions and backmapping
• Lower resolutions 512x512→32x32
• Search for ellipses
• Backmapping : gradually increase the
resolutions and shake the ellipse
Morphological transformation
• Dilation
– A⊕B={cEN| c=a+b for some aA and bB}
• Erosion
– AΘB={xEN| x+bA for every bB}
Procedures
– Data preprocessing
• Image segmentation
– Place white paper on soil
– Manual threshold
• Morphological transformation
• Edge detection
• Thinning
– Hough transform for ellipses
– Focusing
Procedures
•
•
•
•
Original
Threshold + Dilation + Erosion
Edge detection + Thinning
Hough transform for ellipse + Focusing
Object edge image
after thinning.
Resolutions lowered from
512*512 to 32*32 pixels.
Finding possible
ellipses
and sorting them.
Recalculating the
Erasing the duplicated
possibility of
points in the less possible
candidated ellipses .
ellipses.
Increasing resolutions
Choosing the ellipses
from
with threshold.
X*Y to 2X*2Y pixels.
No.
Is it reaching the
final resolutions?
Yes.
Getting the final
list of ellipses.
512x512
128x128
64x64
ellipse
128x128
32x32
64x64
RESULTS
Perimeter regression
Actual perimeter(cm)
30
25
Amaranth
Cabbage
Chinese cabbage
Broccoli
20
15
10
5
0
0
5
10
15
Ellipse perimeter(cm)
20
25
Relationship between actual leaf and ellipse
• Area
• Amaranth
• Cabbage
• Chinese cabbage
• Broccoli
area (cm2)=1.1132*Ellipse area (cm2)+0.0613
(R2=0.954)
area (cm2)=1.1158*Ellipse area (cm2)-0.6975
(R2=0.985)
area (cm2)=1.1386*Ellipse area (cm2)-0.5421
(R2=0.953)
area (cm2)=1.0674*Ellipse area (cm2)-0.068
(R2=0.974)
Relationship between actual leaf and ellipse
• Perimeter
• Amaranth
• Cabbage
• Chinese cabbage
• Broccoli
perimeter (cm)=1.0977*Ellipse perimeter (cm)+1.1233
(R2=0.954)
perimeter (cm)=1.2679*Ellipse perimeter (cm)-0.523
(R2=0.985)
perimeter (cm)=1.1761*Ellipse perimeter (cm)-0.5421
(R2=0.953)
perimeter (cm)=1.2282*Ellipse perimeter (cm)-0.1998
(R2=0.974)
Axial occlusion
Radial occlusion
Area error rate (%)
Cabbage axial occlusion
40.0
20.0
0.0
-20.0
-40.0
-60.0
-80.0
-100.0
-120.0
-140.0
0.0
20.0
40.0
60.0
80.0
Area occlusion ratio (%)
100.0
Perimeter error rate (%)
Cabbage axial occlusion
40.0
20.0
0.0
-20.0
-40.0
-60.0
-80.0
-100.0
-120.0
-140.0
0.0
20.0
40.0
60.0
80.0
Perimeter occlusion ratio (%)
100.0
Area error rate (%)
Cabbage radial occlusion
40.0
20.0
0.0
-20.0
-40.0
-60.0
-80.0
-100.0
-120.0
-140.0
0.0
20.0
40.0
60.0
Area occlusion ratio (%)
80.0
100.0
Cabbage radial occlusion
Perimeter error rate (%)
40.0
20.0
0.0
-20.0
-40.0
-60.0
-80.0
-100.0
-120.0
-140.0
0.0
20.0
40.0
60.0
80.0
Perimeter occlusion ratio (%)
100.0
Broccoli leaf number estimation error rate
Seedling leaves area
120
Chinese cabbage
y = 1.2211x - 3.4339
2
Actual area(cm)
100
80
2
R = 0.910
Cabbage
y = 0.7521x + 3.0748
60
R2 = 0.985
40
Chinese cabbage
Cabbage
20
0
0
40
80
Predicted area(cm2)
120
160
Seedling leaves perimeter
Actual perimeter(cm)
120
100
Chinese cabbage
y = 1.2754x - 4.4678
80
R2 = 0.900
60
Cabbage
y = 0.8112x + 5.7469
R2 = 0.970
40
Chinese cabbage
Cabbage
20
0
0
40
80
Predicted perimeter(cm)
120
CONCLUSIONS
• An image processing algorithm using elliptical
Hough transform is developed to locate seedling
leaves and to estimate leaf area.
• All regressions are highly correlated between
leaf and ellipse area and perimeter.
• Error rate is less than 20% when the occlusion
ratio is under 40% between the actual and
predicted value.
CONCLUSIONS
• When very small object is observed, the initial
processing resolutions should be increased.
• The accuracy to predict the leaf number from
seedling top-view image is above 75%.
• Though the seedling actual leaf area and
perimeter are not the same as the predicted
value, the relationships are highly correlated.
Thank you very much !!
Shake ellipse
Focusing algorithm
• An image size of NxN
• the computational complexity
C=P[16log2N-11]+[1-(t)]24[2t(log2N-t)-log2N]