Computationally Efficient Quantitative Testing of Image Segmentation with a Genetic Algorithm

Quantitative testing of segmentation algorithms implies rigorous testing against ground truth segmentations. Though under-reported in the literature, the performance of a segmentation algorithm depends on the choice of input parameters. The paper reports wide variety both in evaluation time and segmentation results for an example mean-shift algorithm. When testing extends over an algorithmpsilas parameter space, then the search for satisfactory settings has a considerable cost in time. This paper considers the use of a genetic algorithm (GA) to avoid an exhaustive search. As application of the GA drastically reduces search times, the paper investigates how best to apply the GA in terms of initial candidate population, convergence speed, and application of a final polishing round. The GA parameter search forms part of a three-component computation environment aimed at automating the search and reducing the evaluation time. The first component relies on scripted testing and collation of results. The second component transfers to a commodity cluster computer. And the third component applies a genetic algorithm to avoid an exhaustive search.