Active contours driven by Gaussian function and adaptive-scale local correntropy-based K-means clustering for fast image segmentation

摘要

In this paper, we proposed a novel active contour model based on Gaussian function and adaptive-scale local correntropy-based K-means clustering (Gaussian-ASLCK) that is applicable to images with severe intensity inhomogeneity. Firstly, Gaussian function is used to capture the global intensity of images, which boosts the robustness to the initialization. Subsequently, by employing the adaptive scale operator, the scale of the proposed model can be adjusted automatically according to the degree of intensity inhomogeneity. In addition, the regularized term is approximated by a non-local energy, which is constructed based on characteristic functions in the field of heat kernel convolution. Finally, the iterative convolution-thresholding method is designed for minimizing the energy function. This method is simple, efficient and has the optimal complexity of O(Nlog N) per iteration. The experimental results comprehensively validate that the proposed Gaussian-ASLCK model enjoys greater success in images with severe intensity inhomogeneity and complex noise over other state-of-the-art models.