|
  
Next:Model Revision Up:Model Validation Previous:
Raw Pixel Search
Combining results from searching at multiple abstraction levels
The query results from a similarity search are usually ranked according to a similarity index, or score. A higher score indicates increased similarity. The similarity score results from a feature, semantic or
pixel-level search. Multiple constraints are combined using
- Boolean operators, such as AND, OR, NOT;
- Topographical/Spatial operators, such as left, right, contained, within x kilometers, north-north-west of;
- Temporal operators such as before, after;
Figure 6:
An example-based specification of a feature-based search operation.
We use a fuzzy logic approach to combine rankings. Much as in fuzzy set theory, we define a
membership function to measure the similarity of the object Xto the predicate p, and denote it by  . A membership function value of one indicates complete similarity, i.e., equality with respect to the
constraints. Among the possible definitions for the fuzzy union (OR) operation, we adopt the most commonly used one, 
and similarly for the intersection (AND) and the complement (NOT), we adopt the definitions 
The result of a complex query is evaluated according to these membership function composition rules. Note that the definition of the fuzzy AND operation is sensitive to expressions with zero membership. If
any individual membership is zero, the entire composite membership of the phrase is zero. This may be too strong a restriction on the semantics of the query. Alternatively, a weighted approach can be used to
combine results from multiple similarity searches. In this case, the membership function of the AND operation is 
where  is the relative weight of the  predicate of the query phrase.
The following example demonstrates the validation of a model based on texture features. A region of the image has been specified via a rubber band box and supplied as an example for a texture-matching
operator. The specification of the match is shown in Fig. 6. Figure 7 shows the results of the simple query
Find all mixed usage within the currently defined search region. Feature-based objects are stored as square regions of fixed size (in this case 32  32 pixels) which have been pre-extracted from the image. The results represents the best matches to the user-specified example. 
Figure 7: Results of an example-based texture-match query.
Next:Model Revision Up:Model Validation Previous:Raw Pixel Search |
