低秩表示实现鲁棒的CT和MR图像分割

Robust CT and MR Images Segmentation via Low-rank Representation

由于图像中的噪声和复杂性,传统的图像分割方法并不总是能够捕捉到所有细节,即可能会忽略相邻像素的属性或者将图像的不同部分合并在一起.为了充分利用可用信息,利用低秩表示(LRR)和鲁棒主成分分析(RPCA)模型的优点,提出了一种新的图像分割方法,通过模糊c均值(FCM)方法对低秩亲和矩阵进行聚类来获得分割结果.在整个方法中,低秩分量是图像的主要信息,是通过求解RPCA模型获得的,而亲和矩阵表示全局结构,则是通过求解LRR模型获得的.在实验部分,使用计算机断层扫描(CT)图像分割来评估本文方法,结果显示在准确性和鲁棒性方面都有了显著改进.与现有一些算法相比,本文算法对异常值更加鲁棒,并尽可能地保留了图像的细节信息.

It is a challenging task to segment a sequence of images.In fact,much property of these pixels in the images but the classical segmentation approaches are not always able to capture all these details,i.e.the property of adjacent pixels is missed or different parts of the pixels are merged.To take full advantage of the available information,we incorporate the advantage of low-rank representation(LRR)and robust principal component analysis(RPCA)models into our method to segment a sequence of images,finally the segmentation results are obtained by clustering low-rank affinity matrix with fuzzy c-means(FCM)method.In whole method,the low-rank component is obtained by solving RPCA model,and the affinity matrix is obtained by solving LRR model.From the analysis of the properties of two components,we learned that the low-rank component is the primary information of the image;the affinity matrix is the global structure.In the experimental part of the paper,computed tomography(CT)images segmentation are utilized to evaluate our method,which shows significant improvement in both accuracy and robustness.As a result,our algorithm is more robust to the outliers and preserving image details than existing approaches,like FCM clustering,Chan and Vese(CV)and local binary fitting(LBF)approach as much as possible.