CLASS-L Archives

March 2005


Options: Use Proportional Font
Show Text Part by Default
Show All Mail Headers

Message: [<< First] [< Prev] [Next >] [Last >>]
Topic: [<< First] [< Prev] [Next >] [Last >>]
Author: [<< First] [< Prev] [Next >] [Last >>]

Print Reply
"Noordam, Jacco" <[log in to unmask]>
Reply To:
Classification, clustering, and phylogeny estimation
Fri, 4 Mar 2005 09:56:21 +0100
text/plain (26 lines)
Dear cluser and classification Colleagues,

For those of you who are interested in modified fuzzy clustering techniques applied to multispectral images, you can obtain my phd thesis called "Chemometrics in multispectral imaging for quality inspection of postharvest products" at the follwing adress:

a short summary:

This thesis describes different novel chemometric techniques applied to multispectral images for quality inspection on agricultural food products. These images do not only have a huge number of spectral bands which makes training set selection a challenging task, they also contain classes with small defects or abnormalities where objects of these classes are easily missed. For the segmentation and classification of multispectral images the unsupervised Fuzzy C-Means (FCM) clustering algorithm is often used. However, FCM has several known drawbacks which can effect the clustering outcome when applied to multispectral images which contain defects or diseases. One of the drawbacks of FCM, and many unsupervised techniques, is that the spatial information is not used during the classification of such multispectral images. Therefore, two modifications of FCM are presented which combine both spatially and spectrally information into the clustering process to improve image segmentation. Another drawback of FCM is that FCM tends to balance the number of points in each cluster, which results in underestimated defect classes as smaller defect classes are drawn to the larger clusters. A modification of FCM, called cluster insensitive FCM (csi-FCM), is presented in the thesis which overcomes this sensitivity. When the number of spectral bands increases, the huge amount of data in the multispectral images requires computational demands which makes unsupervised segmentation of multispectral images not feasible in most applications. Therefore, a new procedure called Feedback Multivariate Model Selection (FEMOS), is presented which automates the segmentation proces by combining supervised and unsupervised techniques. Chapter 6 presents an application where both multispectral images and RGB color images of French fries with different defects and diseases are evaluated. The explorative analysis of the multispectral images shows that defects are visible in the multispectral images while invisible in the RGB color images and thus for the human eye. The classification results show that the multispectral classification results outperform the RGB color images not only in terms of accuracy but also in terms of yield and purity. Finally, Chapter 7 describes the conclusions and some future aspects of multivariate imaging for agricultural product inspection.

best regards,

J.C. Noordam, phd
Agrotechnology and Food Innovations B.V.
P.O.Box 17,6700 AA Wageningen, the Netherlands
email: [log in to unmask]
tel: +31.317.475139
fax: +31.317.475347
Agrotechnology and Food Innovations participates in GreenVision, the
centre of expertise for image processing in
the agri & food business :