Statistically reinstating method based on the variational framework for fine segmentation

Abstract

In this thesis we suggest a new statistical variational framework and the related curve evolution equation for image segmentation, especially for fine segmentation. It aims to reduce the misclassification error which is inevitable in statistical variational model. To achieve the goal, we introduce the local region-based model which plays a significant role to reduce the misclassification error in the region competition. The local region-based model uses the local distribution of intensities in the local ambiguous region. This model, however, has defects in the sense that it depends on the initial curve for the curve evolution equation and tends to fail to detect the weak edges. Thus by modifying the force term of the curve equation, we make three different region competitions. They are different in which information they use for the design of the force term in the curve evolution equation. Based on these region competitions, we finally propose a novel method which is called the ``Statistically Reinstating Method(SRM)". Furthermore the method adopts the multi-resolution idea to reduce the computational cost. By combining three modified region competitions efficiently we get a fine segmentation. We provide several examples and an application of the method SRM, which is useful in 3D VR content manufacture.