Dynamic Estimation of Curve Evolution in Image Segmentation with CRFs Label Inferring

Typical low level segmentation method like level set method can be explained in maximum a posteriori estimation (MAP) for pixel label. In this paper, CRFs model is introduced in label estimation combined with level set to produce fast low level process and accurate high level inference. The energy term in level set evolution is also extended to contain object spatial factors, gradient is provided as the spatial updating basis, besides the temporal characteristic in curve evolution. Unlike simple CRFs model, a feedback machinery is imported in parameters learning, the reasons lie in the fact that CRFs could has small sample size and its modeling approach is mainly rely on model structure, but image patch is a typical local feature which is not directly applied into. With image patch used in the feedback, the accuracy of learning can be improved. At last, energy function is extended to allow complicated multiple regions competition, the local features is merged in the process.