|
|
|
CorkInspect
is an IST EU project (IST-1999-21088) intending to
support, as a major objective, the early customisation
and validation of a Computer Vision solution for cork
stopper production control (www.cvc.uab.es/corkinspect)
CorkInspect participates in the EUTIST-IMV cluster
(www.spt.fi/eutist)
|
|
|
Cork Stopper Inspection
|
|
`
|
Why?
Automatic
visual inspection of natural products such as food,
cork or wood products is necessary since ‘appearance’
is an important quality factor. It is unthinkable
to have a good and expensive wine bottle with defected
stopper. However, a human inspector’s performance
can vary widely as a result of visual fatigue or other
distractions, and consistent quality can not be guaranteed
to a customer.
What
an automatic computer vision application brings to
a company is a consistent, objective, fast approach
with high percent quality assurance at lower cost.
Moreover, from ecological point of view, the automatic
inspection is necessary to reduce the number of products
classified as non-conformities, minimising waste of
material, that multiplies
its effect when we treat natural products.
Although it does not seem difficult for human beings
to detect different faults in the cork material, it
turns out difficult to precisely formulate the features
of the cork faults due to the porosity of the natural
material. It is difficult even for the cork quality
experts to exactly define all cork features that they
take into account in the process of stopper inspection,
the feature values and ranges in order to define whether
there is a fault in the cork stopper or the stopper
is of poor quality.
There have been different attempts to develop vision
cork inspection systems in the manufacture where the
people working in the manufacture should define the
values and ranges of the image features and elaborate
the decision rules in the process of the stopper inspection.
Given that people in the manufacture work with rather
qualitative than quantitative information to classify
the quality of a stopper, managing such vision cork
inspection systems represent a tedious and time-consuming
task. The problem of the classification of the cork
product in different (in this case, five) quality
groups additionally difficult the problem. This fact
makes difficult for the cork stopper producing company
to define and assure the quality of the products in
front of the providers.
  
   
|
|
|
How?
In order to cope with the problem of subjectivity
in the process of cork inspection and quality classification,
we study different techniques from the fields of Computer
Vision and Pattern Recognition. We propose to apply
statistic algorithms in order to analyze a set of
43 different features of the cork (e.g. number of
cork holes, average stopper gray level, average holes
gray level, holes gray level deviation, length of
largest cork hole, etc), that are considered by the
operators during the cork analysis.
High dimensional data appears in many
pattern recognition problems such as remote sensing,
appearance-based object recognition, text categorization,
etc. A stochastic approach for the classification
of high dimensional data is always a delicate issue.
For linear or quadratic classifiers the number of
training samples depends linearly or quadratically
on the data dimensionality subset of features, an
exhaustive sequential feature selection procedure
is required, so the size of the problem grows combinatorially
on the dimension. Furthermore, the training sample
size needs to increase exponentially in order to effectively
estimate the multivariate densities needed to perform
nonparametric classification.
To avoid the problem of dimensionality,
the most common approach is the implementation of
feature extraction or dimensionality reduction algorithms.
The approach proposed on this project does not seek
dimensionality reduction followed by the implementation
of parametric or nonparametric techniques for density
estimation. Instead, it focuses on the higher level
statistical properties of the data, which is transformed
in such a way that density estimation in the transformed
space is simplified and more accurate.
In order to compare our approach and to
assess which is its performance for the problem of
cork stopper classification, we have implemented and
tested the following methods:
- The most simple classification
technique, the Nearest Neighbor classifier (NN)
, which classifies each cork representation -on
the original space- to the class of the nearest
representation of a stopper from the learning set.
- Principal Component Analysis
(PCA) of the data for dimensionality reduction,
followed by NN classification .
- Linear Discriminant Analysis
(LDA), followed by NN classification.
- Nonparametric Discriminant Analysis
(NDA)
- Maximum Likelihood Classification
(ML) using a Gaussian distribution per class
- Class-Conditional ICA.
|
|
|
|
|
|
|
|
