Optimal bootstrap sampling for fast image segmentation: application to retina image

The authors propose an optimal image sampling model based on the general scheme of bootstrap sampling to get rid of the dependence effect of pixels in real images, and to reduce the segmentation time. Given an original image, a small representative set of pixels is selected randomly. A stochastic model based on the finite normal mixture distribution identification is then used for image segmentation. A local unsupervised segmentation method based on expectation-maximization (EM) family algorithms is then used for parameter estimation, and the maximum likelihood classification (MLC) is adopted for pixel classification. The proposed bootstrap approach is compared with the classical EM family algorithms that make a deterministic sampling pixel after pixel for parameter estimation. The results obtained show that the proposed bootstrap sample selection method gives better results than the classical one in both the quality of the segmented image and the computing time.<>