Automatic scale-dependent adaptation of object models in image analysis (2008)

as part of the DFG bundle project “Abstraction of geoinformation in multi-scale acquisition, management, analysis and visualization"

Cooperation with:

• Institute for Cartography und Geoinformatics, University of Hannover
• Institute for Photogrammetry and Cartography, University FAF Munich
• Institute for Photogrammetry, Institute for Cartography and Geoinformation, Institute of Computer Science, University of Bonn
• Institute of Photogrammetry and Remote Sensing, University of Karlsruhe

 Research group: Automatic image analysis and geographic information

Contact person: Janet Heuwold

Background and Goal

Landscape objects appear differently in varying scales of aerial and satellite images. While in high-resolution images many object details are clearly visible, the same can partly not be identified or distinguished in a lower image resolution. Objects, appearing as areas in high resolution, are recognizable only as points or lines in low scale. For automatic object extraction from aerial and satellite images by means of image analysis, there have to be found appropriate models in the preface. Because of the mentioned background the models differ depending on the resolution of the images to be examined. A repeated creation of image analysis object models for different resolution is redundant, though, as the models for lower resolution can be derived from those ones for higher resolutions (presuming the same image data otherwise).
The aim of this project is to derive automatically image analysis object models for lower resolution from those for higher resolution

Methods and Results

The multi-scale adaptation of the models has been exemplary implemented with roads in suburban areas. The image analysis object models are represented explicitly as Semantic Nets. To guarantee an automatic adaptation process the high-resolution object models have to fulfil certain constraints regarding their composition. The object needs to be described completely and the spatial relations between the object parts have to be clearly defined.
In the automated process of scale-dependent adaptation at first based on the type of object parts in the fine scale a decision is made whether the 1D or 2D scale change analysis method is to be applied: If there are only linear parallel object parts in the object model to be adapted, the 1D scale-dependent adaptation can be used; otherwise the more sophisticated adaptation for 2D objects is to be carried out.
The subsequent process of scale-dependent adaptation consists of three main parts. At first, the object model is decomposed into object parts with similar scale behaviour or parts that need to be analysed together. The influence of adjacent object parts in their scale behaviour needs to be considered here. In the case of influence the scale behaviour of these object parts is analysed jointly in a group. The behaviour of all object parts during scale change is analysed in the scale change models in the next step. The possible occurrence of scale events and the change of appearance of the object parts are determined by a combination of algorithms from differential geometry and image processing. Using analysis-by-synthesis the scale behaviour of the object part (groups) is simulated and analysed automatically. In a last step, the predicted object parts in the target scale are fused back to a complete object model, which is appropriate for the extraction of this object in the lower resolution.

 
Results:
An example for the automatic adaptation of an object model demonstrates the capability of the developed algorithm. The model for a complex road arm with road markings (lines and symbols) in a junction area for aerial images of 0.03m/pix is adapted with the new algorithm to a target resolution of 0.8m/pix

 
Example: Object model for the extraction of a road arm in junction area with direction arrow and stop line in in high spatial resolution of 0.03m/pix

Example: intermediate results analysis-by-synthesis: blob features in high und low resolution superimposed on synthetic images; a) synthetic image high resolution, b) synthetic image low resolution, c) Extrema in high (red) and low resolution (green), d) Blob support regions in high (red) and low resolution (green).

Example: Automatically adapted object model for the extraction of a road arm in junction area with direction arrow and stop line in lower spatial resolution of 0.8m/pix

Conclusions:
A methodology to automatically adapt object models for image analysis to a lower resolution was found. The automatic adaptation of road models was implemented in a sample system.
With the knowledge-based image interpretation system “GeoAIDA” the developed methods were successfully tested on the basis of the automatically adapted object models in different resolutions. The developed adaptation methodology is capable to adapt not only simple models but also more complex object models (even incorporating local context objects), which allow greater flexibility in modelling the objects