Interactive Evolutionary Strategy Based Discovery of Image Segmentation Parameters

Abstract

The symbiosis of human expertise, in terms of creativity and pattern recognition, with evolutionary algorithms for user controlled and directed search is now a rapidly emerging model.

One of the main issues that need to be addressed is the development of techniques to ensure that the power of the evolutionary search is exploited without compromising its efficiency by introducing too much noise in the form of human assessment. Human assessment is likely to have a high component of subjectivity and non-linearity of focus. This implies that in the first instance it is necessary to analyse the nature of the variability of the human assessment. Another important issue that needs to be addressed is ensuring that the evolutionary progress is rapid without compromising the granularity of the search. Rapid convergence is important to the practical applicability of the system and also prevents the process from becoming tedious for the human participant, resulting in loss of concentration.

This paper explores appropriate strategies for the interactive evolution of parameter sets for image segmentation and examines issues relating to reliability of user scores for selection of parents. The nature of user scoring is analysed both in terms of the evolutionary strategy adopted and the temporal progression of the runs. The correlations between number and type of images seen at each generation, the time taken to achieve satisfactory results and the quality of the resulting solutions are analysed in terms of their ability to generalise.