Automatic Adaptation of Segmentation Parameters Applied to Inhomogeneous Objects Detection

Virtually all segmentation methods require parameter tuning, quite a difficult task, mostly performed manually through a troublesome trial-and-error process. To overcome this difficulty, an earlier work describes an automatic parameter adjustment method using Genetic Algorithms (GAs), given an initial set of reference object samples. The method performs well only for homogeneous objects. However, in most real applications, the meaningful image objects are actually non-homogeneous, or rather, an ensemble of usually few homogeneous segments. This work addresses this issue and proposes a supervised GA-based method to automatically adjust the values of segmentation parameters in applications where meaningful objects are inhomogeneous, though formed by an assembly of homogeneous parts. Moreover the work introduces a post-segmentation procedure that merges adjacent segments into single units, which match the geometric form of the interest image objects. Specifically, a metric for detection of polygonal arrangements of segments is proposed herein. Experimental analyses evidence the higher performance of the new method for adjusting segmentation parameters in comparison with the earlier approach. The experiments also attest the ability of the proposed post-segmentation metric to detect polygonal shapes.