5. Machine Vision
Replace human workforce
Non-destructive evaluation
Fast, accurate and reducing cost
V
Vision sensor
5
Image adopted from Shin, J., W. S. Lee, and R. Ehsani. 2012.
Postharvest citrus mass and size estimation using logistic
classification model and watershed algorithm. Biosystems
Engineering 113(1): 42-53
6. Previous Works
Blasco, J., N. Aleixos, and E. Molto. 2006. Computer vision detection
of peel defects in citrus by means of a region oriented segmentation
algorithm. Journal of Food Engineering 81: 535-543.
Region-oriented segmentation algorithm for detecting the most common peel
defects of citrus fruits
Kohno, Y., N. Kondo, M. Iida, M. Kurita, T. Shiigi, Y. Ogawa, T. Kaichi
and S. Okamoto. 2011. Development of a Mobile Grading Machine for
Citrus Fruit, Engineering in Agriculture, Environment and Food
4(1):7-11.
Ripeness and diameter estimation using RGB color images
Cubero, S., N. Aleixos, F. Albert, A. Torregrosa, C. Ortiz, O. Garcia-
Navarrete and J. Blasco. 2013. Optimised computer vision system
for automatic pre-grading of citrus fruit in the field using a mobile
platform. Precision Agric. 15: 80-94
Size and color index estimation using Lab color space
6
11. Objective
11
Develop a fruit inspection system for harvested
citrus fruit on a conveyor belt to identify
healthy, abnormal and small immature fruit for
following year.
13. Conveyor belt system - driven by a tractor
pump
Rotating speed - 48 cm/s
Hardware System
Prototype system for
inspecting citrus on a
conveyor belt
13
14. Imaging System
GigE RGB camera (SXG10C, Baumer)
½” CCD, 1024*1024
Shutter speed: 3 ms
Focal length: 6 mm
FOV: 80 cm by 60 cm
Image acquired in every
1.5 sec
14
Inside of imaging box
15. Image used for experiment
Typical image used for experiment
Total 5 trials (average 18 images per trial)
1 image including 10-30 fruit
15
16. Immature
Types of Citrus Fruit
16
Healthy
Abnormal
Total of 91 images including 617 healthy,
283 abnormal and 237 immature citrus
19. Background Removal
R Threshold = 50
G Threshold = 50
19
Three classes for histogram analysis
Mature: healthy, abnormal
Green: small young fruit for following year
Background: conveyor belt
23. • Ensemble learning
multiple weak learners are trained
and combine their outputs to have
better accuracy.
Random Forest: revised bagging
method with decision tree for
better generalization.
Given a training set of size n, draw a sample
of size n*<n (resampling) and
For i = 1:M
Train classifier Ci, choose a different subset
of features randomly at each split.
end
Final classifier is a majority vote of C1 .. CM
Random Forest Classifier
23
24. Random Forest Classifier
• Features
Color information
- Histogram in R, G, B, H, S, V, Y, Cb, Cr, L, a, b and
gray-scale
- Mean, standard deviation
Size information: area, diameter
Texture
- GLCM: contrast, energy, homogeneity, correlation
- 4 directions: minimum, maximum and average
24
25. Mass estimation
Diameter in cm and in pixels comparison
Regression analysis between mass, diameters
400 oranges (350 mature, 50 immature)
Mass = 0.09*Diameter2
-4.90*Diameter + 80.97
R2 = 0.96
25
27. Image Analysis Result
Trial
No.
Healthy
Oranges
Abnormal
Oranges
Immature Oranges
for Following Year
Leaf &
Conveyor belt
1
Total 88 26 43 248
Correct
Identification (%)
86 (97.7) 22 (84.6) 38 (88.4) 246 (99.2)
2
Total 104 35 22 87
Correct
Identification (%)
97 (93.3) 29 (82.9) 18 (81.8) 78 (89.7)
3
Total 188 91 71 199
Correct
Identification
(%)
176 (93.6) 79 (86.8) 64 (90.1) 181 (91.0)
4
Total 138 76 35 233
Correct
Identification (%)
129 (93.5) 65 (85.5) 32 (91.4) 218 (93.6)
5
Total 99 55 66 180
Correct
Identification (%)
91 (91.9) 42 (76.4) 55 (83.3) 167 (92.7)
Average
Total 123.4 56.6 47.4 189.4
Correct
Identification (%)
115.8
(94.0)
47.4 (83.2) 41.4 (87.0) 178 (93.2)
27
Correctly Identified (%) = 100*Correctly identified object by the algorithm/Ground truth
The number of fruit correctly identified by the algorithm along with
the number of actual fruit by manual counting
28. Mass Estimation Result
28
Mass estimation results of each trial by the algorithm (kg)
Trial No. Actual Mass Estimated Mass R2
RMSE
1 22.4 27.6
2 32.9 33.7
3 50.2 58.13 0.96 4.09
4 43.8 48.0
5 36.4 38.8
Average 37.1 41.2
30. Conclusion
1. A prototype system for fruit inspection of
harvested citrus fruit on a conveyor belt was
developed.
2. Overall classification result was good for
healthy fruit.
3. Irregular patterns of abnormal citrus caused
to lower prediction ability of the classifier.
4. Identifying immature fruit on the conveyor
belt can be used to find effect of mechanical
harvesters for following year crops.
5. Distinguishing healthy and abnormal fruit can
be used for quality evaluation of harvested fruit.
30
Editor's Notes
This presentation consists of four parts, first introduction, second material and method, third result and discussion lastly, conclusion.
This presentation consists of four parts, first introduction, second material and method, third result and discussion lastly, conclusion.
This presentation consists of four parts, first introduction, second material and method, third result and discussion lastly, conclusion.
This presentation consists of four parts, first introduction, second material and method, third result and discussion lastly, conclusion.
This presentation consists of four parts, first introduction, second material and method, third result and discussion lastly, conclusion.