Robust Object Segmentation with Constrained Curve Embedding Potential Field

Abstract

We have earlier introduced an implicit vector field representation for arbitrary number of curves in space, the curve embedding potential field (CEPF), and a general image segmentation strategy based on the detection of the CEPF distortion under the influence of vector-form image data [3]. In this paper, we present an improved CEPF framework which incorporates prior knowledge of the object boundary and has consistent object definition through a region growing process. The embedded implicit curves deform through the image- and model-induced changes of the CEPF, which evidently improves the segmentation accuracy under noisy and broken-edge situations. Further, the closure enforcement and the natural advection on the curves enhance the stability of CEPF evolution and the implementation is straightforward. Robust experimental results on cardiac and brain images are presented.