Information about the Image Matching
What kind of features are used?
We are using three different, partly MPEG-7 conform descriptors, i.e. one local and two global descriptors.
Namly:
- The Edge-Histogram-Descriptor. It is a partly MPEG-7 conform descriptor, that means there is some additional information included in our version of this descriptor. During the extraction process the golbal (not MPEG-7) and semi-global (not MPEG-7) edge features are calculated in addition to the local edge features (MPEG-7). But for the distance measure we are using only the local features. The features are processed with double precision and are quantized afterwards to 8 bit (just for saving harddisk space).
- The Homogeneous-Texture-Descriptor. It is a MPEG-7 conform descriptor, with only few differences in its components to MPEG-7 version. The reason might be found in our grayscale conversion.
- The Color-Layout-Descriptor. It is a MPEG-7 conform descriptor.
All three types of descriptors are stored in a MySQL database and additionaly as xml files on the harddisk.
How is ranking achieved?
All distances are normalized by a offline calculated threshold value so that the simple euclidean distance can work as ranking method.
 |
 |
 |
| histogram of Edge-Histogram distances |
histogram of Homogeneous-Texture distances |
histogram of Color-Layout distances |
The thresholds are displayed as a red bar in each of these three plots.
During the retrieval process the features, i.e. the distances, are combined with respect to the weights the user has chosen to one overall distance.
Functionality of this online retrieval system.
As starting points:
Or