P Systems Inspired Stereo Matching

Abstract

Due to recent hardware advancements, including multi-core and many-core processors, designing parallel versions of sequential algorithms has gained renewed popularity. P system provides a prospective theoretical testbed for designing efficient parallel and parallel-sequential algorithms. In principle, P system is a network of autonomous data processing cells, where each cell transforms its input signals according to its symbol rewriting rules and feeds the results out to its immediate neighbours. This thesis aims to investigate the massive intra- and inter-cell parallelism capabilities of P systems for modelling the symmetric dynamic programming stereo (SDPS) matching algorithm, with particular focus on its inherent parallel computations. We acknowledge that the most common parallelization avenue is by processing each pair of corresponding scanlines of a stereo pair of images independently. Our P system design discovers a new potential avenue for parallel processing that further optimizes the algorithm. Our research of applying Membrane Computing techniques in the field of Computer Vision is positively commended, by our publications reviewers, as the inception of a new research line. Inspired by our P system design of the SDPS algorithm, we experiment with functional and actor-based parallel implementations to leverage the computing power of parallel hardware. Our implementations show that the parallelizable task in our P system design has very fine task granularity, such that the messaging overhead between tasks outweighs the parallelism benefits. We show further results that larger granularity tasks benefit immensely from parallellism.