Orthogonal design of experiments for parameter learning in image segmentation

This paper employs the methods from the design of experiments for supervised parameter learning in image segmentation. We propose to use orthogonal arrays in order to keep the number of experiments small and several algorithms are formulated. Analysis of means is applied to estimate the optimal parameter settings. In addition, a combination of orthogonal arrays and genetic algorithm is used to further improve the performance. The proposed algorithms are experimentally validated based on two segmentation algorithms and the Berkeley image database. A comparison with exhaustive search, an alternating scheme and a Monte-Carlo approach is also provided.

► Orthogonal design based parameter learning helps to reduce computational effort.
► Close-to-optimal solutions are found.
► Orthogonal arrays are suitable for learning parameters of segmentation algorithms.