E cient Parallel Algorithms for the Extraction ofImage

The paper presents algorithms for the construction of features which are invariant with respect to a given nite transformation group. We show that it is in principle possible to construct complete feature sets. The developed methods are general and work for arbitrary nite groups. However, in this paper we focus on translation and rotation invariant features for gray scale images. We investigate possibilities for a parallelization of the algorithms. By analyzing the occurring data dependencies we devise a parallelization strategy which is not only useful for the calculation of invariant features but which can be applied to other typical image processing algorithms as well. The communication demands in the parallel system can be characterized by a set of speciic parameters. According to the values of these parameters it is possible to nd well suited topologies for the communication network. We devise embeddings of these topologies into a xed network structure. Experimental results from an implementation on a 64 processor parallel system are presented.