Geoscience Reference
In-Depth Information
concepts and characterize them. It is possible to apply the generalization approach to
different images (25 images per second) and to the audio track of a video in order to
detect the concepts on each media. Ayache et al. [AYA 06] do not use the same base
of concepts for the two document types: they propose 15 concepts for images and
100 or so for audio documents.
The regrouping approach to generalization consists of gathering several pieces of
information of the same kind. This regrouping does not necessarily lead to
meaningful information as with the identification approach, which does not allow us
to build reference bases such as concepts. For example, the truncation in IR
represents terms with truncated terms (thus, “advantages” and “advantaging” can be
regrouped into “advantag”). Pham et al. [PHA 07] propose the application of this
approach to images by splitting the image with a regular grid to obtain a set of cells
(called patches). An image can thus be considered a bag of patches (just as a text is
considered a bag of terms). Pham et al. call these patches visterms (visual terms) by
analogy with the terms contained in texts (which they call texterms). When splitting
an image in visterms using a regular grid, the same visterm can appear several times;
thus, its raw frequency determines its weight just like for the terms. Pham et al. also
propose a segmentation of the images in regions, which consists of performing a
splitting guided by the content (colors, for example). Nevertheless, this approach
does not allow the semantic identification of the constructed regions. Standardization and decision support systems
Mainly used in multicriteria decision support, standardization consists, just like
normalization, of bringing the score of the different criteria along the same interval
of results ([0;1], in general). Nevertheless, the main difference with normalization is
that standardization is applied to non-numeric evaluations [MAL 03, LAB 03]. For
example, each criterion is evaluated following three categories (strong interest,
average interest and weak interest) in [MAL 03] and following six categories (very
weak, weak, average, strong, very strong and exceptional) in [LAB 03]. Moreover,
standardization demands that, for each criterion, the categories be compared in pairs.
In [MAL 03], each pair is labeled from 1 to 9 by the user, which indicates whether
the two categories are of equal importance (1) or one of them is a lot more important
than the other (9). Once this pair-comparison matrix has been established, the system
generates the corresponding standardized numeric values [SAA 80, MAL 03]. An
example of a standardization process with an added comparison phase of category
pairs is illustrated in [PAL 10a]. It is a study in the automotive domain. The
categories corresponding to the color criteria are “ideal” for blue and red,
“acceptable” for green and orange and “borderline” for white and black. The
comparison phase of these category pairs determines the values of 1 for “ideal”, 0.75
for “acceptable” and 0.3 for “borderline”: the green and orange cars will thus have a
weight of 0.75 for the color criteria.
Search WWH ::

Custom Search