Stimulation Initiative for European Neural Applications
Esprit Project 9811
Automatic Sorting of Pot Plants
The problem
Actual topics in plant nursing are quality monitoring and improvement of
the process efficiency in the greenhouse. A high quality product with a
low cost price is essential to handle increasing competition. A consequence
is the need of further mechanization and automation. Sorting of plants
and flowers on quality is one of the few tasks that still is mainly performed
by humans. Labour cost for sorting and product handling are a substantial
part of the cost price. An additional disadvantage of human grading is
the subjectivity of the judgement. For instance, when grading on size humans
have the tendency to compare the plants in a group and to select the extreme
plants and put them in the class of small and big plants. The grading criteria
can be changed completely when a group with another size distribution is
examined. Considering the fact that quality of plants is determined by
numerous plant features, like size, height, color, shape, symmetry, etc.
it is not surprising that humans have problems to maintain the same objective
decision criteria. In most situations several people sort plants on quality
and then it is practically impossible to hold on to the requested constant
grading criteria.
Neural network application
ATO-DLO
has developed a flexible and universal grading system for pot plants. The
system consists of a color camera, an image processing system and a classification
system. The system is completed with conveyer belts. The maximum capacity
of the system is about 10000 plants per hour. The core of the classification
system is either a conventional statistical classifier or a neural network.
The selection of the classifier depends on the complexity of the application.
Conventional statistical classifiers have only limited classification capabilities,
but they are easier in use. Neural networks, on the other hand, proved
to be superior for more complex classifications problems. In both cases,
the system is trained by showing product examples to the camera. Currently,
the system has been installed at several greenhouses in the Netherlands.
Benefits
Learning techniques - statistical classifiers or neural networks - allow
the user to store knowledge into the computer without the need to formulate
this knowledge in explicit rules - a difficult and laborious job. With
learning techniques, the system can easily be adjusted to meet new requirements.
The automated visual grading system guarantees a constant grading criteria
during operation while it saves on labour costs.
Generalization
The key feature of this application is the use of a learning classifier
which analyzes images from a computer vision system. Automated visual inspection
has possible applications across the whole of the manufacturing industry
where large numbers of products need to be checked for quality. In particular
for complex classification problems where conventional statistical classifiers
fail neural networks might be able to provide a solution.
Contact person
Ir. A. Timmermans. ATO-DLO, P.O.
Box 17, 6700 AA Wageningen, The Netherlands. Tel. +31 3174 75271, Fax +31
3174 12260.
References for further reading
A. Timmermans. Automatic Sorting of Pot Plants with a Neural Network Classifier.
In:
Neural Networks: Artificial Intelligence
and Industrial Applications. Proceedings of the 3rd Annual SNN Symposium
on Neural Networks, Nijmegen, The Netherlands, 14-15 September 1995, Nijmegen.
B. Kappen and S. Gielen (eds). Springer-Verlag, London.
![[ ]](../logo_small.gif){kind=logo}
SIENA