自动分割图像新方法研究

Research of new automatic image segmentation method

提出一种采用非参数Dirichlet过程混合模型实现图像自动分割的算法。该方法在图像分割时不需要对分类数进行初始化,具有在分割过程中自动获得图像分类数的特点。模型中使用有控制参数的随机变量来代替聚类数,通过调整参数来指定聚类数的范围。使用该算法对具有高噪声的自然图像和临床磁共振图像进行分割实验,并与其他分割算法进行比较。实验结果显示本算法抗噪声性能强,且可以抑制磁共振图像分割过程中的偏场效应。准确度分析显示,图像分割结果的Dice相似性系数均高于90%,表明提出的新算法具有很高的精确性和鲁棒性。

This paper applied nonparametric Dirichlet process mixtures model algorithm to segment images automatically, which could obtain the segmentation class numbers automatically during the process without initialization. The number of clusters as an input constant was substituted by a random variable with a control parameter, which specified a level of cluster resolution by adjusting the parameter. The algorithm was used to segment noisy natural images and magnetic resonance images with biasing field. The experiment results show that the algorithm has anti-noise performance and can restrain the biasing field effect of MR images well. The Dice similarity coefficients are all above 90% ,which show that the proposed method is robust and accurate.