Search

Image databases that allow retrieval of data by referring to the image contents need an image analysis module. In this paper we will describe the features of ‘I-see’, a knowledge-based software environment that we have designed specifically for image analysis. We will assess the usefulness of this tool for the development of interpretation modules for image databases. Two important research topics are investigated. First, precompilation of the image contents is necessary for efficient retrieval. Second, query by image contents must allow access to objects and situations represented in the images. Our approach will be to look at these database functionalities in general image analysis terms. In this sense, precompilation will be considered as a special case of image interpretation, query specification as a form of model description while query retrieval will be viewed in terms of model instantiation and selection. Although the ‘I-see’ tool already contains features that can be used as image database parts, the paper will consider the tool as a research platform rather than as an image database module.

Knowledge based image analysis is a combination of digital signal processing and symbolic reasoning. In this paper, we will look at some problems connected to the symbolic reasoning approach to image interpretation and see how an iconic representation can help to solve some of them. We will show that many of the features and problems connected with both symbolic and iconic representation are complementary.

Road networks extracted from multi-temporal SPOT images of the same scene are matched to collect evidence for individual road segments. The Dempster-Shafer theory is applied to find a degree of confirmation for a road segment in one network based on its corresponding lines in the other networks.

The paper describes a software environment in development that supports the implementation of image understanding applications. The described environment, called I-see, is implemented on top of the object oriented KEE ∗ system and focuses on the iconical representation and exploration of visual data. The system offers an interface between the high-level symbolic reasoning mechanisms of KEE and the raw image data and iconically represented image models or segmentation results.

A software environment that supports the implementation of image understanding applications is described. The environment, called I-see, is implemented on top of the object-oriented KEE system and focuses on the iconic representation and exploration of visual data. The system offers an interface between the high-level symbolic reasoning mechanisms of KEE and the raw image data and iconically represented image models or segmentation results. An example problem of interpreting satellite images shows the use of this iconically represented knowledge in the form of a terrain elevation image. In this specific example, the iconic representation could be used directly without the need for a complementary symbolic description. The iconic and symbolic representation schemes can be used to supplement each other and can be related by the use of label images. >

Currently both statistical and neural classifiers are being used for the classification of multispectral satellite imagery. Because both classifier types are being used as 'black boxes' and because they are values based on different mathematical models the reasons for their different performance levels are not well understood. The authors have used visualization of class decision boundaries in feature space as a means to gain insight into the classification processes. >

The paper explores the usefulness and applicability of knowledge-based image interpretation. By limiting the analysis to ‘restricted’ scenes, a bottom-up strategy has been developed to improve a primal image segmentation. Two case studies are discussed: the first deals with medical X-rays, the second with satellite images (SPOT). In both projects, generic geometrical knowledge is encoded in the format of production rules. The results obtained so far are encouraging and are already of practical use. Ways to extend the knowledge bases by more specific domain knowledge are mentioned.

Automated delineation of linear cultural structures can help to improve the classification of remotely-sensed images. This topic also provides an excellent testbed for knowledge-based computer vision research. In this paper, a road network extraction system, useful on SPOT satellite images, is described. By applying semantic model-fitting operators, an initial spatial segmentation is obtained. To analyse the resulting primal road network, declaratively stated knowledge about generic appearances of roads and crossroads on SPOT images is used. As a result a more detailed description of the network can be obtained. Advantages and drawbacks of the use of existing knowledge base tools for this computer vision task are reported.

Work in progress on knowledge-based blood vessel segmentation of subtraction angiograms is reported. The proposed approach is centered around different modular stages which are linked through the concept of a multi-layered image representation. At the bottom level classical image processing techniques are used to extract edges, center lines and bars. At the intermediate stage, low-level geometrical knowledge has been encoded to construct blood vessel segments. The higher levels will contain general blood vessel knowledge and application domain-dependent knowledge in order to create a final representation of the blood vessels. Although the system is only partly implemented, a result already of practical use will be discussed.