目的 针对LCK(local correntropy-based K-means)模型对初始轮廓敏感的问题,提出了新的基于全局和局部相关熵的GLCK(global and local correntropy-based K-means)动态组合模型。方法 首先将相关熵准则引入到CV(Chan-Vese)模型中,得到新的基于全局相关熵的GCK(global correntropy-based K-means)模型。然后,结合LCK模型,提出GLCK组合模型,并给出一种动态组合算法来优化GLCK模型。该模型分两步来完成分割:第1步,用GCK模型分割出目标的大致轮廓;第2步,将上一步得到的轮廓作为LCK模型的初始轮廓,对图像进行精确分割。结果 主观上,对自然图像和人工合成图像进行分割,并同LCK模型、LBF模型以及CV模型进行对比,结果表明本文所提模型的鲁棒性比上述模型都要好;客观上,对BSD库中的两幅自然图像进行分割,并采用Jaccard相似性比率进行定量分析,准确率分别为91.37%和89.12%。结论 本文算法主要适用于分割含有未知噪声及灰度分布不均匀的医学图像及结构简单的自然图像,并且分割结果对初始轮廓具有鲁棒性。

Objective The local correntropy-based k-means (LCK) model can segment an image that contains unknown noise and has an uneven gray distribution. However, the segmentation result is sensitive to the initial contour. To solve this problem, a new dynamic model based on global correntropy-based k-means (GCK) and LCK is presented. Method The dynamic model is a combination of two models. A new algorithm, i.e., GCK, is proposed by introducing correntropy to the coefficient of variation (CV) model and improving the CV model. A global and local correntropy-based k-means (GLCK) model is then proposed by combining GCK and LCK dynamically to retain each method's advantages. The GLCK model is not a simple linear combination of the two models. The model implements two steps to complete segmentation. First, the GCK model isutilized to segment an image and obtain the general outline of the image. Second, the image with the initial contour as segmentation results of GCK is segmented finely by LCK. To improve segmentation accuracy, a dynamic combination algorithm is designed by controlling the time when the GCK model transforms into the LCK model automatically. Result The segmentation result of the proposed method is compared to that of three other similar segmentation methods, namely, LCK, local binary fitting, and CV models, on natural and synthetic images. Results showed that the proposed model is more robust than the three other models. By segmenting two natural images on the BSD library and using the Jaccard similarity ratio for quantitative analysis, accuracy rates of 91.37% and 89.12% are obtained. Conclusion The proposed algorithm can effectively segment medical images and the simple structure of natural images with unknown noise and an uneven gray distribution; the result is robust to the initial outline.