Current Issue Cover
空间可变有限混合模型

申小虎,吕导中,万荣春(江苏警官学院刑事科学技术系, 南京 210031)

摘 要
目的 有限混合模型是一种无监督学习方法,它被广泛的应用到数据分类任务中。然而,在图像分割过程中,由于有限混合模型没有引入邻域像素间的空间关系,导致了图像分割结果对噪声非常敏感。为了增强有限混合模型的抗噪性,提出一种新的空间可变有限混合模型。方法 该模型通过在像素的先验分布中引入一种新的空间关系来降低噪声对图像分割结果的干扰。在构建空间关系的过程中,利用形态学膨胀原理将空间邻域内特征值出现的概率而不是特征值本身进行膨胀操作,然后通过根据具有最大概率的分类标记在高斯混合模型迭代地计算过程中进行局部像素空间平滑,从而起到抑制噪声干扰的作用。结果 本文实验包含了人工合成图像和医学CT图像的图像分割实验。在人工合成图像分割实验中,对人工合成图像添加了不同程度的噪声来测试本文模型和对比模型对噪声抑制能力的高低;对医学CT图像进行图像分割实验,以是比较本文模型与对比模型之间在实际图像分割中的效果。结论 实验数据显示,本文提出的模型在噪声抑制能力上,图像分割精度和计算效率上均有更优的性能。
关键词
Spatially variant finite mixture model

Shen Xiaohu,Lyu Daozhong,Wan Rongchun(Department of Forensic Science Technology of Jiangsu Police Institute, Nanjing 210031, China)

Abstract
Objective The finite mixture model (FMM) is an unsupervised learning method that is widely applied to data classification tasks, particularly image segmentation. The Gaussian mixture model is a successful example of FMM used in image segmentation. However, segmentation result is sensitive to noise because the spatial relationship among neighboring pixels is not considered. To solve this problem, spatially variant FMM (SVFMM) and its improvements have been proposed by incorporating spatial constraints into the prior distribution of each pixel. These improvements have been demonstrated to be capable of noise suppression. The spatial constraint of SVFMM has been widely studied and improved. To enhance the robustness of FMM against noise, a new SVFMM is proposed in this study. Method The proposed model based on the concept of morphological dilation considers the existence of a spatial relation in the posteriori probability distribution of the pixel neighborhood to reduce the interference of noise in image segmentation result. The proposed model is also introduced to prior probability distribution. Spatial relationships are incorporated into prior distribution for spatial smoothness by redesigning morphological dilation. The concept of morphological dilation is adopted to increase the probability of the features of a pixel in the statistic instead of the feature value itself. The neighboring pixels are smoothened iteratively by the label with the highest probability in the neighborhood. To maximize the likelihood function, gradient descent technology instead of the expectation-maximization algorithm is employed to estimate the parameters of the proposed model. Result The proposed model is implemented via MATLAB. Experimental data include synthetic and medical computed tomography (CT) images. Synthetic images are interrupted by different degrees of noise to test the robustness against noise of the proposed model. Medical CT images are used to analyze effectiveness in real applications. Experiments on image segmentation show that the proposed model exhibits considerable noise suppression effect and computational efficiency. Compared with the existing SVFMM improvements presented in literature, the proposed model uses less parameters in estimation, is easier to implement, and has lower computational cost. Conclusion The proposed model is superior to compared models in terms of robustness against noise, segmentation accuracy, and computational efficiency. The computational efficiency of the proposed model is better than that of most SVFMMs with spatial constraints in terms of image segmentation performance. The result for the CT images shows that this research can provide valuable help in analyzing similar criminal cases. In the field of criminal investigation, accurate extraction of a segmented region is a prerequisite to analyze images related to crime.
Keywords
QQ在线


关注微信