Goal of FEX:

FEX (Feature EXtraction) allows to extract simultaneously all geometric primitives images basically have, and which can be useful to describe the image structure in different application fields. Additionally, FEX also extracts the topological neighborhood relations between the features. Features and relations provide a symbolic description of the image, represented as a planar attributed graph, the ''Feature Adjacency Graph'' FAG.

In detail FEX extracts:

Points:
- Junctions (intersections of lines)
- Dots (isolated, circular symmetric, small point-type regions)
Lines:
- Step Edges (contour lines, i.e. common boundary of two regions)
- Bar Edges like drawings (elongated, line-type regions, maybe isolated or open, i.e. not connected to other lines)
Blobs:
- homogeneous image areas.
Topological neighborhood relations between the features

Some Characteristics of FEX

FEX needs no specific knowledge about the scene, thus FEX can be applied for a broad range of application fields or image types. The image model assumes just the following prerequisites: the image area can be segmented into homogeneous areas with well defined boundaries and where the boundaries can be approximated by straight line segments.
To make FEX also for non-specialists usable we reduced the control file to 4 interpretable parameters which must be predefined by the user in dependency of the image type and resolution, the image quality and the application (the object models). These 4 parameters are: 1 resolution scale (also for noise reduction), 2 scales for lines and points (integration kernels), and a significance level for internal statistical tests. The definition of thresholds etc. is not necessary.
Before feature location takes place we extract areas of interest. This allows to select feature type dependent location procedures which are appropriate and fit best to the local image function.
After the location within the areas the areas are set to background. This leads to disjunct features, but also non-connected features. Thus, at local node structures we may have extracted points of type junction and also the lines, but the points and the line endpoints are never connected (incident) after FEX. We therefore analyze local point structures in a second procedure called FAGANA to close the structures (not documented yet, but results are also shown below).

The FEX control file provides the users to select or combine any feature types and relations which are appropriate and useful in the distinct application. Thus, it is possible to extract only the features or relations being of interest. This holds for the feature types (points(P), lines (L)and/or blobs(B)) and their subclasses, the neighborhood relations between the different feature types (7 combinations: all (P-L-B), only between points (P-P), lines(L-L), or blobs (B-B), or without points(L-B), lines(P-B), blobs(P-L)), but also for the feature representation types: I.e., it is also possible to define the edge pixel chains as ''line feature'' instead of the chain approximation, i.e. the polygons. The FAG is stored in an ASCII output file in a well-defined format, called ''Global Exchange Format'' (cf. FEX home page). These output files produced by FEX build the input data for several other projects at our, and also some extern institutes.

FAGANA:

FAGANA is a grouping procedure which is appplied after FEX. The goals are:

Improvement of the segmentation results after FEX, e.g. the closing of point structures and contour lines which are after FEX still open due to the ''background''-strategy of FEX
Extraction or grouping of more complex structures or aggregates (structural pyramid).

The most important charcteristic of this procedure is, that we do not integrate additional knowledge about the image scene than already used for FEX. Other projects at the ipb (e.g. building reconstruction, cf. Felicitas Lang) involves building models for the FAG analysis after FEX, but these projects are not treated here.

The neighborhood relations stored in the FAG directly allow to analyze the image structure on a next level of complexity. We group the features to more complex structures which is simply based on a topological model of an ideal image partitioning. Without more knowledge as already used in FEX we can use the FAG to group on a next level three type of more complex structures:

Nodes: Points with their adjacent lines
Wings: Lines with their adjacent blobs at both sides
Cells: Blobs with their adjacent (surrounding) lines and nodes

The FAG can not only used to build up a structural pyramid, but also to find errors or inconsistencies of the description due to image noise, low contrasts, etc. In a complete and error-free image partitioning several relationships would not exist, e.g.: We would have no topological neighbored points without a connecting line, no neighbored lines without a connecting point and no neighbored blobs without a common boundary line in between.

When analysing real images it is mostly the case that we do extract such neighborhood relations. But these inconsistencies in the FAG after FEX can be directly use to initiate further analysis at these points of attention. E.g. we add in FAGANA lines between to blobs if the contour is locally splitted and we add also junction points as intersections of lines, if a junction point was not extractable by the point extraction procedure in FEX. The features, which are added within FAGANA to make the image partitioning locally complete are called VIRTUAL lines and VIRTUAL points.

The demo consists of a sequence of images in GIF format, which can be down-loaded.

Principle: A table of what can be extracted, and an overview of input and output data

Principle: Flow Diagram of FEX, including a short explanation of the procedure.

Synthetic Image: Here we summarized an example in more detail, i.~e. including some intermediate results (cf. the FEX Flow Diagram shown in the image before).

For the next examples we show for several input images (upper left) each the extracted features (upper right), the image of the exoskeleton (lower left), and also the results after FAGANA in the lower right image. The visualization of the rsults of FAGANA does not contain the extracted blobs (here, we want to focus only on the lines and points structures).

1. Example: Synthetic TESTPATTERN Image (this is identical to the 1. Example, but now also shows the results after FAGANA)

2. Example: Building front in a terrestrial image

3. Example: Building in aerial image, large scale

4. Example: Building in aerial image, small scale

5. Example: Building (cf. 4. Example) in aerial image, large scale