Ontologies for Object Recognition, 10-R9733Printer Friendly Version
Inclusive Dates: 07/01/07 11/20/08
Background - Object recognition is the method by which objects in an image are classified. Object recognition is a task, which in the past has been very difficult to automate, despite being a task that is easily mastered by people. Traditional computer vision approaches to object recognition break down an image into smaller images or objects, with each being represented by a set of features. Pattern recognition methods are then used to compare the unknown object features to a set of known objects. A common problem in pattern recognition is when multiple known objects are matched to the unknown object. In this case, because of a limited feature set, the unknown object does not uniquely map to a single known object. For some object recognition applications this limitation was not an issue because the number of known objects could be kept small. In addition, the image and objects of interests were well controlled, allowing a high probability of accurate recognition by traditional methods. However, future applications of object recognition will require methods that can operate in a dynamic environment with imperfect information. Object recognition methods will have to recognize a large number of objects that may not be uniquely defined by their features.
Approach - To achieve the goal of recognizing a large number of objects, contextual information and feature data must both be considered. For example, a book may not be uniquely defined by its features alone, but if a book were shown on a bookshelf in an image, that additional contextual information could be used to classify the object as a book. Software ontologies are ideal for incorporating context into object classification because they allow the efficient storage and reasoning of common sense knowledge that is required to define and use contextual information. The incorporation of software ontologies into object recognition is expected to vastly improve object recognition over traditional methods. To test this, simulated objects and their corresponding features have been classified by the new ontology method and by the traditional pattern recognition method. A simulation has been used to accelerate the development of the ontology based object recognition. The simulation data is based on a real life problem of satellite imagery analysis, so that the results are credible, even though based on a simulation.
Accomplishments - The research program is now in the final testing phase. The developed algorithms are being evaluated using a variety of satellite images. The algorithms are currently being compared against traditional recognition techniques to quantify the improved performance achieved with context.