Towards object-based retrieval for image libraries

Object discovery is one of the steps to effective content-based image retrieval. If a vision system has the ability to identify potential objects in an image, it enables the user to phrase queries in terms of objects, and it enables the computer to search the image library for similar objects. The paper describes a physics-based approach to object discovery. The approach grows seed regions to include neighboring regions when they are deemed coherent according to a suite of physics-based and heuristic measures. Using this approach, the computer is able to extract coherent surfaces from an image library and use the resulting object images to execute searches. Initial results indicate higher accuracy and precision for object-based image retrieval when compared to traditional histogram-based retrieval. The object-based method is able to find and retrieve images that are considered poor matches by other methods