Parallel imperative and functional approaches to visual scene labelling

摘要

Low and intermediate image processing produce a description of a visual scene in terms of constituent features and their interrelationships. We have developed both parallel imperative and functional forms of an algorithm for the interpretation of segmented scene data, which effectively match a dynamic data structure representing the scene against a database of preformed models. These models represent components which may exist within the scene, either in complete view or partially obscured. For evaluation and comparison of the two approaches, the algorithms have been implemented in occam and mi. respectively, and tested on images of industrial components. Currently, these tests have been restricted to a single procesor but the algorithms are designed for general purpose multiple instruction multiple data (MIMD) machines.