Parallel Evaluation of Hierarchical Image Databases

Abstract

This paper presents a hierarchical image transformation and efficient parallel algorithms for its evaluation. This transformation maps image structures onto code trees of different height, depending on the size of the structure. Thereby, important structures are effectively separated from the background. The inherent parallelism of such a hierarchical image transformation is outlined. The algorithms are domain independent and were successfully used for workpiece recognition and for traffic sign detection. A communication module for farmer-worker applications that supports specialized processors, like frame grabbers or display units, as well as the parallel recognition process is illustrated in detail. The implementation is done on a 9-node transputer image processing system. The functionality from grabbing an image and low-level filtering to transformation and high-level symbolic pattern analysis is integrated. Runtimes for the analysis of traffic scenes of less than half a second were achieved. Illustrated examples and experimental results are discussed.