A unified object-oriented toolkit for discrete contextual computer vision

This paper describes a new general-purpose contextual architecture which provides a unified framework for efficiently combining all types and levels of context in discrete computer vision applications, by organising the multidimensional search space in best-first order along each dimension. It then implements an efficient 'lazy evaluation' algorithm, which searches from the most probable vertex outwards, and guarantees to find solutions in absolute best-first order. The architecture has been designed and built as a C++ class library, and utilised within a demonstrator which implements full contextual constraints for optical character recognition of hand-printed postal addresses. Preliminary evaluation of the demonstrator suggests the system has the potential to achieve genuinely remarkable performance compared with previous context systems: its memory requirements are an order of magnitude less than an equivalent trie-based dictionary; its search speed is at least an order of magnitude faster than the trie, and actually get faster as the dictionary size increases; and its error rate is virtually zero, even when an OCR system with appalling performance is simulated. Using this architecture it appears to be possible to build real-time solutions to large-scale contextual vision problems which have previously been beyond the bounds of computational feasibility.