Articles
NEURAL NETWORKS USED FOR CLASSIFICATION OF POTTED PLANTS
Article number
562_11
Pages
109 – 115
Language
English
Abstract
In the framework of the EC-funded AIR-project Objective plant quality measurement by image processing feature extraction and decision system software is developed.
The decision system is a neural network (NN) which is used for sorting (classification) of plants.
The NN sorts plants, based on some relevant plant features, extracted from digital plant images.
To be able to sort plants correctly the NN needs to be trained first.
Weights in the NN are adapted so that for a batch of plants, the inputs of the NN (plant features) generate the desired output, which is the class the plant is assigned to by a human expert.
In general, there is no ranking order in the classes and therefore classes are nominal data.
The NN has an output neurone for each of the classes the expert distinguished.
For detection of the NN adapting to the inconsistencies in sorting by the expert, the available plants are randomly divided in a trainset and a testset.
The NN is used for classifying Ficus benjamina in marketable stage.
It is trained to imitate each of six human experts who sorted the 300 available plants.
By using up to three features as input for the NN, an 89% performance was reached.
The NN sorted 89% of the plants the same as the human expert.
For the expert whose sorting was most difficult to imitate a 71% performance was reached.
These results are acceptable considering the fact that the human experts introduce inconsistencies when sorting plants.
The NN is used to create uniform growth groups in Ficus benjamina plants at a half-grown stages.
The NN was trained on half-grown plants from a reference experiment.
Results are varied, probably because the half-grown plants from the reference experiment do not resemble the actual half-grown Ficus benjamina plants.
The decision system is a neural network (NN) which is used for sorting (classification) of plants.
The NN sorts plants, based on some relevant plant features, extracted from digital plant images.
To be able to sort plants correctly the NN needs to be trained first.
Weights in the NN are adapted so that for a batch of plants, the inputs of the NN (plant features) generate the desired output, which is the class the plant is assigned to by a human expert.
In general, there is no ranking order in the classes and therefore classes are nominal data.
The NN has an output neurone for each of the classes the expert distinguished.
For detection of the NN adapting to the inconsistencies in sorting by the expert, the available plants are randomly divided in a trainset and a testset.
The NN is used for classifying Ficus benjamina in marketable stage.
It is trained to imitate each of six human experts who sorted the 300 available plants.
By using up to three features as input for the NN, an 89% performance was reached.
The NN sorted 89% of the plants the same as the human expert.
For the expert whose sorting was most difficult to imitate a 71% performance was reached.
These results are acceptable considering the fact that the human experts introduce inconsistencies when sorting plants.
The NN is used to create uniform growth groups in Ficus benjamina plants at a half-grown stages.
The NN was trained on half-grown plants from a reference experiment.
Results are varied, probably because the half-grown plants from the reference experiment do not resemble the actual half-grown Ficus benjamina plants.
Authors
C.J.H.M. van Kaam
Keywords
decision system, neural networks, inconsistency, Ficus benjamina
Online Articles (50)