Homogeneity Testing for Unlabeled Data: A Performance Evaluation

Abstract

In this paper, we address the problem of testing homogeneity for unlabeled pixels observed in a subimage. Homogeneity testing is an essential component in split-and-merge segmentation algorithm. Two types of homogeneity tests are involved: tests for labeled data when deciding on merges between regions and tests for unlabeled data when deciding whether to split a region. In our study, we focus on images that are modeled as a mosaic of uniform regions corrupted by additive Gaussian noise. Using this model, we present a statistical analysis on the performance of two commonly used approaches for testing homogeneity of unlabeled data based on the region/subregion similarity and the data dispersion, respectively. We also propose and evaluate a new hierarchical homogeneity testing scheme for unlabeled data. The most important finding of this study is that the tests based on region/subregion similarity have a low power on average of detecting inhomogeneity in unlabeled data.