Binary segmentation, a problem of extracting foreground objects from the background, often arises in medical imaging and document processing. Popular existing solutions include expectation maximization (EM) algorithm, Otsu thresholding, K-sigma thresholding, and the recently proposed generalized principal component analysis (GPCA). We apply these algorithms to segmentation of noisy images with small foreground objects. Such images often arise in change detection applications such as functional magnetic resonance imaging (fMRI). In our experiments none of the algorithms performed sufficient well when the total size of foreground regions was much smaller than the size of the background region. We propose a novel algorithm, called sGPCA, that can robustly estimate the intensity of small foreground objects in the presence of noise. The intensity estimate obtained can be used to determine an optimal threshold value or to initialize EM and Markov random field (MRF) based segmentation algorithms.
