Scale-space representation of remote sensing images using an object-oriented approach

摘要

Today's high resolution remotely sensed images (<1m) pose several challenges which require solutions that go beyond the traditional spectral based methodologies. With the rapid increase in the level of detail present in these images, there is also an increase in the complexity. To deal with this complexity a consistent framework and image representation is needed. An object-based scale-space representation is proposed. Principles of objectbased design are explained and the application of these principles to image regions is introduced. Given an input image, the scale-tree is automatically constructed using low-level information, starting with single pixels (as objects) and ending with the root node indicating the complete image. The scale-tree is a hierarchical structure where each level in the hierarchy differs from the next in the size of the objects/regions present at that level. Hence, the scale-tree reflects the scale-space breakdown of the image. From another point of view the scaletree can also be seen as a collection of multiple segmentations with varying level of detail going from fine to coarse. Synthetic and real high resolution satellite images were used to evaluate our image representation. The goal of the proposed representation is to facilitate applications such as target/anomaly detection, image region classification and change detection.