Transformation-invariant indexing and machine discovery for computer vision

Future computer vision systems must have the ability to discover new object models. This problem can be addressed by relational concept formation systems, which structure a stream of observations into a taxonomy of discovered concepts. This paper presents a representation for images which is invariant under arbitrary groups of transformations. The discovered models, also being invariant, can be used as indices for 3D images. The methodology is illustrated on a small problem in molecular scene analysis, where discovered models, invariant under Euclidean transformations, are efficiently recognized in a cluttered molecular scene.