A multiobjective metaheuristic approach for morphological filters on many-core architectures

Abstract

Mathematical Morphology (MM) is a set-theoretic technique for the analysis of geometrical structures. It provides a powerful tool for image processing, but is hampered by significant computational requirements. These requirements can be substantially reduced by decomposing complex operators into sequences of simpler operators, at the cost of degradation of the quality of the results. This decomposition also directly translates to streaming task graphs, a programming model that maps well to the kind of systolic architectures typically associated with many-core systems. There is however a trade-off between mappings that implement high-quality filters and mappings that offer high performance in many-core systems. The approach presented in this paper exploits a multi-objective evolutionary algorithm as a design-time tool to investigate trade-offs between the quality of the MM decomposition and computational performance. The evolutionary process performs an analysis of filter quality vs computational performance and generates a set of task graphs and mappings that represent different trade-offs between the two objectives. It then outputs a Pareto front of mapping solutions, allowing the designer to select an implementation that matches application-specific requirements. The performance of the tool is benchmarked on a morphological filter for the detection of features in a high-resolution PCB image.