12/12/2002 09:00 AM
Electrical Engineering
Many image segmentation techniques create a partition containing more regions than is practical for the particular application. These cases require the application of region count reduction procedures in order to reduce the partition\'s region count to a more desirable number. In this paper, we present a region count reduction algorithm that represents a significant departure from other published algorithms designed for the same task. Our novel approach is centered around a versatile, multiphase, region-merge algorithm that targets specific image features for merging on a per-phase basis. This new approach allows fast execution of the region-merge algorithm since only specific regions participate in each region-merge phase. Moreover, region descriptors are based on values that allow for constant-time updating after two regions are merged. The concept of aggregate region size is introduced and used as a region descriptor. Unlike most published region count reduction techniques, our new algorithm does not require the calculation or monitoring of region merging termination criteria. The results show the new region-merge algorithm consistently produces viable partitions over a set of test images. Moreover, these results show that the lack of region merging termination criteria is not detrimental to the visual quality of the final partition. The multiphase approach is able to adapt to various types of images despite the relative simplicity of the region descriptors used. The results also indicate a modest reduction in required processing time when compared to a standard approach to region merging.
