| Summary: | 碩士 === 國立中興大學 === 農業機械工程學系 === 83 === The purpose of the thesis is to combine vision processing
techniques with neural network to establish a grading system of
carrots. The system used modified moment to characterize carrot
contour and used classifier to identify forked carrot. Then the
ystem utilized the cross-gradient operator and the
4-connectivity labeling method to detect cracks. Those carrots
which were not forked or cracked were further graded into three
classes. In the grading processes, the crack-following method
was employed to find the contour pixels and the Hotelling
transform method was used to obtain the major axis. Then, the
symmetry, diameter difference, curvature, and compactness of
the carrot image were acquired according to the relationships
between the major axis and the contour. The green area of the
carrot image was calculated by adding all the green pixels of
the image. Whether a pixel was classified as green or red was
based on the chromaticity diagram. The size of the green area
was used as the color describing factor. A three-layer error
back-propagation neural network employing the shape and color
factors as the input parameters was used to simulate the manual
grading of carrots. To search for the best grading model, the
types and number of the input parameters and the number of the
nodes of the hidden layer were changed. The result of this
study revealed that the neural network grading system could
identify forks approximately at 96%, cracks at 92% and the
grading accuracy at 81% in comparison with manual grading.
|
|---|
