Print

发布时间: 2020-02-16
摘要点击次数:
全文下载次数:
DOI: 10.11834/jig.190337
2020 | Volume 25 | Number 2




    遥感图像处理    




  <<上一篇 




  下一篇>> 





空间约束层次加权Gamma混合模型的SAR图像分割
expand article info 石雪, 李玉, 赵泉华
辽宁工程技术大学测绘与地理科学学院遥感科学与应用研究所, 阜新 123000

摘要

目的 合成孔径雷达(SAR)图像中像素强度统计分布呈现出复杂的特性,而传统混合模型难以建模非对称、重尾或多峰等特性的分布。为了准确建模SAR图像统计分布并得到高精度分割结果,本文提出一种利用空间约束层次加权Gamma混合模型(HWGaMM)的SAR图像分割算法。方法 采用Gamma分布的加权和定义混合组份;考虑到同质区域内像素强度的差异性和异质区域间像素强度的相似性,采用混合组份加权和定义HWGaMM结构。采用马尔可夫随机场(MRF)建模像素空间位置关系,利用中心像素及其邻域像素的后验概率定义混合权重以将像素邻域关系引入HWGaMM,构建空间约束HWGaMM,以降低SAR图像内固有斑点噪声的影响。提出算法结合M-H(Metropolis-Hastings)和期望最大化算法(EM)求解模型参数,以实现快速SAR图像分割。该求解方法避免了M-H算法效率低的缺陷,同时克服了EM算法难以求解Gamma分布中形状参数的问题。结果 采用3种传统混合模型分割算法作为对比算法进行分割实验。拟合直方图结果表明本文算法具有准确建模复杂统计分布的能力。在分割精度上,本文算法比基于高斯混合模型(GMM)、Gamma分布和Gamma混合模型(GaMM)分割算法分别提高33%,29%和9%。在分割时间上,本文算法虽然比GMM算法多64 s,但与基于Gamma分布和GaMM算法相比较分别快600 s和420 s。因此,本文算法比传统M-H算法的分割效率有很大的提高。结论 提出一种空间约束HWGaMM的SAR图像分割算法,实验结果表明提出的HWGaMM算法具有准确建模复杂统计分布的能力,且具有较高的精度和效率。

关键词

SAR图像分割; 层次化加权Gamma混合模型; 马尔可夫随机场; 期望最大化法; M-H算法

Hierarchically weighted Gamma mixture model with spatial constraint for SAR image segmentation
expand article info Shi Xue, Li Yu, Zhao Quanhua
Institute for Remote Sensing Science and Application, School of Geomatics, Liaoning Technical University, Fuxin 123000, China
Supported by: National Natural Science Foundation of China (41301479, 41271435); Natural Science Foundation of Liaoning Province, China(2015020090)

Abstract

Objective The development of synthetic aperture radar (SAR) technology has resulted in the generation of high-resolution SAR images under all weather conditions and time periods. SAR images are widely used in many fields, such as disaster monitoring and ocean science. SAR image segmentation is a crucial step in image processing. The statistical model-based SAR segmentation algorithm is popular for its statistical distribution of homogeneous regions in SAR images with specific regularity. However, the statistical distribution of pixel intensities in high-resolution SAR images can be asymmetric, heavy-tailed, or multi-modal. Traditional mixture models use the weighted sum of components to model the statistical distribution of pixel intensities in SAR image segmentation. The components of mixture models are defined by probability density functions to mainly model the statistical distribution of homogeneous regions. The components can be Gaussian, student's, or Gamma distribution in the Gaussian mixture model (GMM), student's mixture model, and gamma mixture model (GaMM), respectively. However, these components fail to model the complicated distribution of pixel intensities in SAR images. To address the problem, this study proposes a SAR image segmentation algorithm that is based on a hierarchically weighted Gamma mixture model (HWGaMM) with spatial constraint. Method A mixture model is defined by the weighted sum of its components to model the statistical distribution of pixel intensities. Its components are usually defined by the probability distribution, which results in the difficulty of modeling the complicated distribution of homogeneous regions in SAR images. To accurately model the asymmetric, heavy-tailed, or multi-modal distribution of pixel intensities, the proposed algorithm uses the HWGaMM to model the statistical distribution of pixel intensities in SAR images. The component of the HWGaMM is defined by the weighted sum of Gamma distributions, which represent the element used to model the statistical distribution of local homogeneous regions. As a result of the differences in pixel intensities in the same region and the similarities of pixel intensities in different regions for high-resolution SAR images, the HWGaMM is defined by the weighted sum of the components. The hierarchy of the HWGaMM can be expressed as follows. The basic layer is the element, i.e., Gamma distribution, which is used to model the statistical distribution of local homogeneous regions. The second layer is the component, which is the weighted sum of elements to mainly model the statistical distribution of homogeneous regions. The top layer is the HWGaMM, which is the weighted sum of components to model the statistical distribution of SAR images. The spatial relation of pixels is modeled by a Markov random field to reduce the influence of image noise. The spatial relation of pixels is introduced to the HWGaMM by defining the weight of components by the posterior probabilities of the pixels and neighboring pixels. Such introduction can improve the robustness of the proposed algorithm and prevent the increase in the complexity of model parameter estimation. In this work, SAR image segmentation is realized by estimating the model parameters through the combination of the Metropolis-Hastings (M-H) algorithm and expectation maximization (EM) algorithm. The traditional M-H algorithm usually suffers from poor efficiency because of its sampling for every model parameter in each iteration. The EM algorithm cannot easily estimate the shape parameter of a Gamma distribution because the shape parameter is included in the gamma function. To address such problem, the proposed algorithm uses the M-H algorithm in simulating the posterior distribution of the shape parameter and the EM algorithm in estimating the scale parameter and element weight. The method of parameter estimation overcomes the drawback of the EM algorithm and achieves higher efficiency than the M-H algorithm. Result Segmentation experiments are carried out on simulated and real SAR images, and the results are analyzed qualitatively and quantitatively to verify the feasibility and effectiveness of the proposed algorithm. The proposed algorithm is compared with the GMM-based, Gamma distribution-based, and GaMM-based segmentation algorithms to highlight its advantages. The results of the fitting histograms reveal that the HWGaMM can accurately model the complicated distribution of pixel intensities. The segmentation accuracy can be calculated from the confusion matrix to quantitatively evaluate the proposed algorithm. The segmentation accuracies of the proposed algorithm are 33%, 29%, and 9% higher than those of the GMM-based, Gamma distribution-based, and GaMM-based segmentation algorithms, respectively. The segmentation time of the proposed algorithm is 64 s faster than that of GMM-based segmentation algorithm but is 600 s and 420 s slower than that of the gamma distribution-and GaMM-based segmentation algorithms, respectively. Conclusion This work proposes an SAR image segmentation algorithm that is based on a spatially constrained HWGaMM. The proposed HWGaMM can model the complicated distribution of pixel intensities. The proposed segmentation algorithm also has higher accuracy than other relevant methods. Although the efficiency of the proposed algorithm is lower than that of the EM-based segmentation algorithm, it is much higher than that of the M-H-based segmentation algorithm.

Key words

synthetic aperture radar (SAR) image segmentation; hierarchically weighted Gamma mixture model (HWGaMM); Markov random field (MRF); expectation maximization (EM); Metropolis-Hastings (M-H) algorithm

0 引言

图像分割是图像处理过程中的关键步骤之一,是将图像划分为互不重叠的同质区域。随着合成孔径雷达(SAR)技术的不断发展,使其在各领域得到了广泛应用。但SAR图像内固有的相干斑噪声给图像分割带来了难题和挑战。学者们提出了基于阈值(张海涛和李雅男,2015Ji等,2016Yu等,2017谌华等,2019)、聚类(赵雪梅等,2014Ji和Wang,2014)、统计(夏梦琴等,2015Wang和Shi,2018唐德可等,2019)等SAR图像分割方法。SAR图像各同质区域内像素强度统计分布呈现出某种规律,因此基于统计模型的SAR图像分割受到了广泛关注(张金静等,2016)。

学者们提出了许多基于统计模型的SAR图像分割方法(Wang等,2015Xu等,2017Song等,2017)。其中一种分割方法是,假设待分割图像各区域像素强度服从独立同一的概率密度函数,以此建立图像的统计模型(Gao,2010)。如王玉等人(2016)假设像素强度服从同一独立的Gamma分布,根据贝叶斯定理构建图像分割模型,并设计M-H(Metropolis-Hastings)算法模拟分割模型实现图像分割。在M-H算法模拟分割模型中,设计了3个更新操作,需要在迭代过程中依次执行各更新操作直到算法收敛,这大大降低了该算法的效率。张英海等人(2016)利用Gamma分布建模SAR图像同质区域内像素强度统计分布,且各分布相互独立,并采用期望条件最大化(ECM)算法估计模型参数并实现SAR图像分割。在采用ECM算法求解参数中,由于Gamma分布的形状参数包含于Gamma函数中,这导致ECM算法难以求得其解析解。为此,采用Newton-Raphson方法近似计算形状参数的数值解,该迭代方法的计算量较大,降低了该算法的效率。上述基于单一分布的分割算法中,通过设定Gamma分布的形状和尺度参数,该分布可呈现出非对称和右侧重尾的形态。因此,该分布适用于建模SAR图像统计分布。但上述分割方法没有考虑同质区域内统计分布的复杂性,这导致其对图像噪声敏感且分割精度低。为了准确建模SAR图像的统计分布,Zhao等人(2017)采用Gamma混合模型(GaMM)建模像素强度统计分布,利用马尔可夫随机场(MRF)建模像素邻域关系,并通过最大化边缘分布实现SAR图像分割。GaMM利用多个Gamma分布加权和建模SAR图像统计分布,比基于单一分布的分割算法更具有鲁棒性。除此以外,由于高斯分布结构简单,易于实现,高斯混合模型(GMM)在建模遥感图像上受到了广泛关注(赵泉华等,2017)。但高斯分布为对称分布,难以准确建模SAR图像的复杂统计分布。Zhang等人(2014)采用学生t混合模型(SMM)建立SAR图像统计分布模型,学生t分布仍为对称分布,其结构较高斯分布复杂,但鲁棒性优于高斯分布。上述分割算法难以建模SAR图像内像素强度呈现的非对称、重尾或多峰等复杂统计分布。

为了准确建模SAR图像统计分布并实现精确SAR图像分割,石雪等人(2018)提出了层次Gamma混合模型的高分辨率SAR图像分割算法,该算法具有准确建模SAR图像像素强度统计分布的能力,但该算法假定类别之间像素强度统计分布相互独立,没有考虑到同质区域内像素强度的差异性和异质区域间像素强度的相似性,且该算法采用M-H算法模拟分割模型实现图像分割,这需要进行大量的采样以优化各模型参数,大大降低了图像分割的效率。

为此,本文提出一种空间约束层次加权Gamma混合模型(HWGaMM)的SAR图像分割算法。HWGaMM的混合组份由多个Gamma分布加权和定义,考虑到同质区域内像素强度的差异性和异质区域间像素强度的相似性,采用多个混合组份的加权和定义HWGaMM。由于SAR图像中固有斑点噪声的影响,提出算法采用MRF建模像素的空间位置关系,利用各像素及其邻域像素的后验概率定义混合权重,以此构建出空间约束HWGaMM。为了实现高效且精确的SAR图像分割,提出算法结合M-H和EM算法求解模型参数。该参数求解方法克服了M-H算法需要对所有参数逐个进行大量采样所导致效率低的缺陷,同时避免了EM算法难以求解形状参数解析式的问题。为了验证提出算法的可行性和有效性,对仿真和实测SAR图像进行分割实验,并对实验结果做定性和定量分析。实验结果表明提出的HWGaMM算法可准确建模像素强度的复杂统计分布,且提出算法可获得高精度的分割结果。在分割效率上,虽然提出算法的效率比基于EM的分割算法低,但与基于M-H的分割算法相比较,其效率有很大的提高,因此,提出算法具有较高的分割效率。

1 提出算法

1.1 空间约束HWGaMM的建立

给定一幅SAR强度图像$\mathit{\boldsymbol{x = }}\left({{x_i}, \;i = 1, 2, \cdots, n} \right)$,其中$i$为像素索引,${x_i}$为像素$i$的强度,$n$为总像素数。为了准确建立像素强度统计分布模型,采用HWGaMM建模SAR图像的统计模型,则${x_i}$概率密度分布为

$ p\left({{x_i}\left| {{\mathit{\boldsymbol{ \boldsymbol{\varPsi} }}_i}} \right.} \right) = \sum\limits_{l = 1}^k {{\pi _{li}}} p\left({{x_i}\left| {{\mathit{\boldsymbol{ \boldsymbol{\varOmega} }}_l}} \right.} \right) $ (1)

式中,${\mathit{\boldsymbol{ \boldsymbol{\varPsi} }}_i} = \left\{ {{\mathit{\boldsymbol{\pi }}_i}, \mathit{\boldsymbol{ \boldsymbol{\varOmega} }}} \right\}$为HWGaMM的模型参数集,${\mathit{\boldsymbol{\pi }}_i} = \left\{ {{\pi _{li}};l = 1, \cdots, k} \right\}$为混合权重,$\mathit{\boldsymbol{ \boldsymbol{\varOmega} }} = \left\{ {{\mathit{\boldsymbol{ \boldsymbol{\varOmega} }}_l};l = 1, \cdots, k} \right\}$为混合组份的参数集,$l$为混合组份(类别)索引,$k$为总类别数,${{\pi _{li}}}$表征像素$i$隶属于类别$l$的先验概率,满足条件0 < ${{\pi _{li}}}$ < 1和$\sum\limits_{l = 1}^k {{\pi _{li}} = 1} $。混合组份为

$ p\left({{x_i}\left| {{\mathit{\boldsymbol{ \boldsymbol{\varOmega} }}_l}} \right.} \right) = \sum\limits_{j = 1}^m {{{w_{lij}}}} {G_a}\left({{x_i}\left| {{\theta _{lj}}} \right.} \right) $ (2)

式中,${\mathit{\boldsymbol{ \boldsymbol{\varOmega} }}_l} = \left\{ {{\mathit{\boldsymbol{w}}_l}, {\mathit{\boldsymbol{\theta }}_l}} \right\}$为第$l$个混合组份的参数集,${w_l}$ = {${{w_{lij}}}$; $i$ = 1, …, $n$, $j$=1, …, $m$}为分量权重,${\theta _l}$ = {${\theta _{lj}}$; $j$ = 1, …, $m$}为分量参数集,$j$为分量索引,$m$>1为总分量数,${{w_{lij}}}$表征像素$i$隶属于类别$l$中第$j$个分量的概率,其满足条件0 < ${{w_{lij}}}$ < 1和$\sum\limits_{j = 1}^m {{w_{lij}}} = 1$,分量${G_a}\left({{x_i}|{\theta _{lj}}} \right)$定义为Gamma分布,即

$ {G_a}\left({{x_i}\left| {{\theta _{lj}}} \right.} \right) = \frac{{x_i^{{\alpha _{lj}} - 1}}}{{\mathit{\Gamma }\left({{\alpha _{lj}}} \right)\beta _l^{{\alpha _{lj}}}j}}\exp \left({ - \frac{{{x_i}}}{{{\beta _{lj}}}}} \right) $ (3)

式中,${\theta _{lj}} = \left\{ {{\alpha _{lj}}, {\beta _{lj}}} \right\}$${{\alpha _{lj}}}$${{\beta _{lj}}}$分别为形状和尺度参数,$\mathit{\Gamma }\left(\cdot \right)$为Gamma函数。

在统计学上假设各像素强度之间相互独立,则SAR图像的概率分布可表示为

$ \begin{array}{*{20}{c}} {p\left({\mathit{\boldsymbol{x}}\left| {{\mathit{\boldsymbol{ \boldsymbol{\varPsi} }}_i}} \right.} \right) = \prod\limits_{i = 1}^n {p\left({{x_i}\left| {{\mathit{\boldsymbol{ \boldsymbol{\varPsi} }}_i}} \right.} \right)} = }\\ {\prod\limits_{i = 1}^n {\left[ {\sum\limits_{l = 1}^k {{\pi _{li}}} \sum\limits_{j = 1}^m {{w_{lij}}} \frac{{x_i^{{\alpha _{ij}} - 1}}}{{\mathit{\Gamma }\left({{\alpha _{lj}}} \right)\beta _l^{{\alpha _{lj}}}j}}\exp \left({ - \frac{{{x_i}}}{{{\beta _{lj}}}}} \right)} \right]} } \end{array} $ (4)

式中,模型参数集可进一步写为$\mathit{\boldsymbol{ \boldsymbol{\varPsi} }} = \left\{ {\mathit{\boldsymbol{\pi }}, \mathit{\boldsymbol{w, \alpha, \beta }}} \right\}$,其中,$\mathit{\boldsymbol{\pi }} = \left\{ {{\mathit{\boldsymbol{\pi }}_i};i = 1, \cdots, n} \right\}$$\mathit{\boldsymbol{w}} = \left\{ {{w_l};l = 1, \cdots, k} \right\}$$\mathit{\boldsymbol{\alpha }} = \left\{ {{\alpha _{lj}};l = 1, \cdots, k, j = 1, \cdots, m} \right\}$$\mathit{\boldsymbol{\beta }} = \left\{ {{\beta _{lj}};l = 1, \cdots, k, j = 1, \cdots, m} \right\}$

混合权重${{\pi _{li}}}$表示像素$i$隶属于类别$l$的先验概率,根据贝叶斯定理(赵泉华等,2017)可构建像素$i$隶属于类别$l$的后验概率为

$ {z_{li}} = \frac{{{\pi _{li}}\sum\limits_{j = 1}^m {{w_{lij}}{G_a}\left({{x_i}\left| {{\theta _{lj}}} \right.} \right)} }}{{\sum\limits_{l' = 1}^k {{\pi _{l'i}}} \sum\limits_{j = 1}^m {{w_{l'ij}}{G_a}\left({{x_i}\left| {{\theta _{l'j}}} \right.} \right)} }} $ (5)

式中,分量权重${{w_{lij}}}$表示像素$i$隶属于类别$l$中第$j$个分量的先验概率,根据贝叶斯定理可构建像素$i$隶属于类别$l$中第$j$个分量的后验概率为

$ {y_{lij}} = \frac{{{w_{lij}}{G_a}\left({{x_i}\left| {{\theta _{lj}}} \right.} \right)}}{{\sum\limits_{j' = 1}^m {{w_{lij'}}} {G_a}\left({{x_i}\left| {{\theta _{lj'}}} \right.} \right)}} $ (6)

为了克服SAR图像内固有相干斑噪声对图像分割造成的影响,利用MRF建模各像素的空间位置关系,以构建空间约束HWGaMM。本文算法以像素及其邻域像素的后验概率定义混合权重,以将像素邻域关系引入HWGaMM来提高算法的抗噪性。其公式为

$ {\pi _{li}} = \frac{{\exp \left({\eta \sum\limits_{i' \in {\mathit{\boldsymbol{N}}_i}} {{z_{li'}}} } \right)}}{{\sum\limits_{l' = 1}^k {\exp } \left({\eta \sum\limits_{i' \in {\mathit{\boldsymbol{N}}_i}} {{z_{l'i'}}} } \right)}} $ (7)

式中,${i'}$为像素$i$的邻域像素索引,${\mathit{\boldsymbol{N}}_i}$为邻域像素索引集合,$\eta $为邻域像素对中心像素的平滑作用系数。

1.2 模型参数求解

将SAR图像分割问题转化为模型参数求解问题。从式(4)可看出形状参数${{\alpha _{lj}}}$存在于$\mathit{\Gamma }\left({{\alpha _{lj}}} \right)$中,这导致难以求得其解析解。为此,提出算法采用M-H算法(Wang等,2015)模拟形状参数的后验分布,以实现形状参数的求解。但由于M-H算法需大量采样以优化参数,这导致其求解效率低。为此,提出算法采用EM算法求解具有解析解的参数${{\beta _{lj}}}$${{w_{lij}}}$

1) 形状参数求解。定义${{\alpha _{lj}}}$的先验分布,假设其服从均值为${\mu _\alpha }$,标准差为${\sigma _a}$的高斯分布,并假设其组份和分量之间相互独立,则形状参数的先验分布为

$ \begin{array}{*{20}{c}} {p\left(\mathit{\boldsymbol{\alpha }} \right) = \prod\limits_{l = 1}^k {\prod\limits_{j = 1}^m p } \left({{\alpha _{lj}}} \right) = }\\ {\prod\limits_{l = 1}^k {\prod\limits_{j = 1}^m {\frac{1}{{\sqrt {2{\rm{ \mathit{ π} }}\sigma _\alpha ^2} }}} } \exp \left({ - \frac{{{{\left({{\alpha _{lj}} - {\mu _\alpha }} \right)}^2}}}{{2\sigma _\alpha ^2}}} \right)} \end{array} $ (8)

然后,根据贝叶斯定理,结合HWGaMM和形状参数先验分布构建其后验分布为

$ \begin{array}{*{20}{c}} {p\left({\mathit{\boldsymbol{\alpha }}\left| \mathit{\boldsymbol{x}} \right.} \right) \propto p\left({\mathit{\boldsymbol{x}}\left| \mathit{\boldsymbol{ \boldsymbol{\varPsi} }} \right.} \right)p\left(\mathit{\boldsymbol{\alpha }} \right) = }\\ {\prod\limits_{i = 1}^n {\left[ {\sum\limits_{l = 1}^k {{\pi _{li}}} \sum\limits_{j = 1}^m {{w_{lij}}} {G_a}\left({{x_i}\left| {{\theta _{lj}}} \right.} \right)} \right]} \times }\\ {\prod\limits_{l = 1}^k {\prod\limits_{j = 1}^m {\frac{1}{{\sqrt {2{\rm{ \mathit{ π} }}\sigma _\alpha ^2} }}} } \exp \left({ - \frac{{{{\left({{\alpha _{lj}} - {\mu _\alpha }} \right)}^2}}}{{2\sigma _\alpha ^2}}} \right)} \end{array} $ (9)

最后,设计更新形状参数操作,具体过程如下:随机选择类别$l$和分量$j$,所对应的${{\alpha _{lj}}}$为待更新的形状参数,以${{\alpha _{lj}}}$为均值、$\varepsilon $为标准差的高斯分布生成候选形状参数$\alpha _{ij}^*$,其接受率如

$ \begin{array}{*{20}{c}} {a\left({\mathit{\boldsymbol{\alpha }}, {\mathit{\boldsymbol{\alpha }}^ * }} \right) = \min \left\{ {1, \frac{{\exp \left({ - \frac{1}{{2\sigma _\alpha ^2}}{{\left({\alpha _{lj}^* - {\mu _\alpha }} \right)}^2}} \right)}}{{\exp \left({ - \frac{1}{{2\sigma _\alpha ^2}}{{\left({{\alpha _{lj}} - {\mu _\alpha }} \right)}^2}} \right)}} \times } \right.}\\ {\left. {\frac{{\prod\limits_{i = 1}^n {\sum\limits_{l = 1}^k {{\pi _{li}}} } \sum\limits_{j = 1}^m {{w_{lij}}} {G_a}\left({{x_i}\left| {\alpha _{lj}^*, {\beta _{lj}}} \right.} \right)}}{{\prod\limits_{i = 1}^n {\sum\limits_{l = 1}^k {{\pi _{li}}} } \sum\limits_{j = 1}^m {{w_{lij}}} {G_a}\left({{x_i}\left| {{\alpha _{lj}}, {\beta _{lj}}} \right.} \right)}}} \right\}} \end{array} $ (10)

若接受率大于1则接受$\alpha _{ij}^*$,否则保持${{\alpha _{lj}}}$不变。

2) 尺度参数和分量权重求解。对于可求得解析解的尺度参数和分量权重,采用EM算法实现(赵泉华等,2017)。在E步,将最大化似然函数转化为最大化似然函数期望。对似然函数取对数,并取期望作为新的目标函数,即

$ \begin{array}{*{20}{c}} {J = \sum\limits_{i = 1}^n {\sum\limits_{l = 1}^k {{z_{li}}} } \left[ {\log {\pi _{li}} + \sum\limits_{j = 1}^m {{y_{lij}}} \left({\log {w_{lij}} + } \right.} \right.}\\ {\left. {\left. {\left({{\alpha _{lj}} - 1} \right)\log {x_i} - \log \mathit{\Gamma }\left({{\alpha _{lj}}} \right) + {\alpha _{lj}}\log {\beta _{lj}} - \frac{{{x_i}}}{{{\beta _{lj}}}}} \right)} \right]} \end{array} $ (11)

在M步,利用上式对${{\beta _{lj}}}$求偏导,并令其为0,可得到尺度参数的解析式为

$ {\beta _{lj}} = \frac{{\sum\limits_{i = 1}^n {{z_{li}}} {y_{lij}}{x_i}}}{{\sum\limits_{i = 1}^n {{z_{li}}} {y_{lij}}{\alpha _{lj}}}} $ (12)

由于${{w_{lij}}}$需满足其约束条件,因此采用拉格朗日乘数法定义其约束目标函数为

$ {J_w} = J + {\lambda _w}\left({\sum\limits_{j = 1}^m {{w_{lij}}} - 1} \right) $ (13)

式中,${\lambda _w}$为分量权重的拉格朗日系数。利用上式对${{w_{lij}}}$求偏导,并令其为0,得分量权重为

$ {w_{lij}} = \frac{{{y_{lij}}}}{{\sum\limits_{j' = 1}^m {{y_{lij'}}} }} $ (14)

本文空间约束HWGaMM算法流程如图 1所示。

图 1 提出算法流程图
Fig. 1 Flowchart of the proposed algorithm

2 实验结果与讨论

为了验证空间约束HWGaMM分割算法的可行性和有效性,在Intel Core i5-3470 CPU@ 3.20 GHz,8 GB内存,MATLAB R2016a环境下,采用提出算法对仿真和实测SAR图像进行分割实验,并对实验结果进行定性和定量分析。利用GMM分割算法(赵泉华等,2017)、Gamma分割算法(王玉等,2016)、GaMM分割算法(Zhao等,2017)作为对比算法。实验中HWGaMM算法的常数设定如下,组份数$m$设为2,邻域作用参数$\eta $设为0.5,迭代次数$T$设为1 000。

2.1 仿真SAR图像分割

图 2(a)为模板图像,其中Ⅰ—Ⅳ为各同质区域的标号。根据表 1的参数生成仿真SAR图像,如图 2(b)所示。仿真SAR图像中各区域内的像素强度是从两组参数不同的Gamma分布中随机抽取,根据权重值从2个分布中抽取的像素数分别为各区域像素数(64×64)的60%和40%,且所抽取的像素随机分布在各区域内。因此,该仿真图像各同质区域的像素强度统计分布较为复杂,以此可验证HWGaMM算法的统计建模能力和分割的有效性。

图 2 模板和仿真图像
Fig. 2 Template image and simulated image ((a) template image; (b) simulated image)

表 1 仿真SAR图像参数设置
Table 1 Setting parameters of simulated image

下载CSV
组分权重区域Ⅰ区域Ⅱ区域Ⅲ区域Ⅳ
αβαβαβαβ
10.642515205360
20.443810404450

图 3为采用GMM算法、Gamma算法、GaMM算法和HWGaMM算法对仿真SAR图像进行分割的结果。图 3(a)中GMM算法难以将各区域区分开,各区域均存在不同程度的误分割像素。图 3(b)中Gamma算法可将黑色区域分割开,但其他区域均存在误分割斑块。图 3(c)中GaMM算法的分割结果中同样存在较多的误分割像素。图 3(d)中HWGaMM算法可将各区域分割开,且分割结果中仅存在极少的误分割像素。综上,空间约束HWGaMM算法的分割结果在视觉上明显优于对比算法。

图 3 仿真SAR图像分割结果
Fig. 3 Segmentation results of simulated images ((a)GMM; (b)Gamma; (c)GaMM; (d)HWGaMM)

为了定量评价各分割算法,利用模板图像分别与各算法的分割结果进行比较计算混淆矩阵,根据混淆矩阵分别计算用户、产品(制图)和总精度及kappa值,见表 2。以总精度为例进行说明,GMM算法的总精度最低,为66.16%,Gamma算法的总精度为70.62%,GaMM算法的总精度为90.82%,空间约束HWGaMM算法的总精度为99.61%,明显高于对比算法, 其他精度也高于对比算法。因此认为,提出算法可获得高精度分割结果。

表 2 各分割结果的精度评价结果
Table 2 Accuracy evaluation of segmentation results 

下载CSV
/%
算法精度区域Ⅰ区域Ⅱ区域Ⅲ区域Ⅳ
GMM用户98.9561.4745.2761.48
产品94.4361.2847.5661.38
总精度66.16
kappa0.59
Gamma用户99.4278.5244.3262.57
产品99.9870.2746.4465.82
总精度70.62
kappa0.64
GaMM用户98.7778.7397.1691.81
产品98.3994.9788.6781.25
总精度90.82
kappa0.88
HWGaMM用户100.0099.7199.9398.80
产品99.9599.4499.3299.73
总精度99.61
kappa0.99

为了验证HWGaMM的建模能力,利用GMM算法、Gamma算法和HWGaMM算法的参数估计值,绘制仿真图像各区域的直方图拟合结果,如图 4所示。由于Gamma算法和GaMM算法均采用Gamma分布建模同质区域的统计分布,且GaMM算法为模糊聚类方法,因此该实验仅考虑Gamma算法的建模结果。图 4(a)中区域Ⅰ直方图较为陡峭且右侧重尾,Gamma算法和HWGaMM算法均可对其准确拟合,因此对应该区域的分割结果较为精确。但由于高斯分布的对称性,GMM算法在右侧尾部拟合较差,因此对应区域分割结果存在误分割现象。图 4(b)中区域Ⅱ直方图平缓且右侧重尾,GMM算法明显难以拟合该类直方图,而Gamma算法和HWGaMM算法均可对其准确拟合。图 4(c)中区域Ⅲ直方图呈现多峰和右侧重尾特性,仅HWGaMM算法可准确拟合该直方图。图 4(d)中区域Ⅳ直方图较为平坦且覆盖的强度范围广,GMM算法难以拟合该类直方图,Gamma算法可近似拟合,而HWGaMM算法的拟合准确。综上,HWGaMM算法具有准确建模各复杂统计分布的能力。

图 4 各区域直方图拟合结果
Fig. 4 Results of fitting histograms ((a) region Ⅰ; (b)region Ⅱ; (c)region Ⅲ; (d)region Ⅳ)

为了定量评价直方图拟合结果,计算各算法拟合曲线的拟合误差,即$e = \sum\limits_{s = 1}^{256} {{{\left({{h_i} - {{\hat h}_i}} \right)}^2}} $,其中${{{\hat h}_i}}$为拟合曲线中强度$i$估计概率值,${h_i}$为直方图中强度$i$真实概率值,见表 3。从表 3中可以看出,GMM算法的拟合误差较大,Gamma算法的拟合误差明显低于GMM算法,而提出算法的拟合误差优于Gamma算法,尤其是对区域Ⅲ的非对称且双峰分布的拟合,因此提出算法具有较高的拟合精度。

表 3 直方图拟合曲线的定量评价
Table 3 Results of quantitatively evaluating the fitting curve

下载CSV
算法拟合误差$e$/10-3
区域Ⅰ区域Ⅱ区域Ⅲ区域Ⅳ
GMM3.62.86.346.8
Gamma0.420.250.9645.3
HWGaMM0.320.400.2844.6

对提出算法的复杂度进行分析,用o(${N_z}$)表示类别后验概率的复杂度,o(${N_y}$)表示分量后验概率复杂度,o(${N_w}$)和o(${{N_\pi }}$)表示权重参数的复杂度,均与图像大小相关,o(2${{N_\theta }}$)表示分量参数的复杂度,上述复杂度均与图像大小和迭代次数$T$相关,因此算法的总复杂度为$T$×o(${N_z}$+${N_y}$+${N_w}$+${{N_\pi }}$+2${{N_\theta }}$)。表 4列出各算法的分割时间。GMM算法采用EM算法求解模型参数,因此具有最高的分割效率。Gamma算法采用M-H算法需进行多次采样更新模型参数,导致该算法效率最低。GaMM算法需要多次迭代,导致该算法效率较低。提出算法结合了EM和M-H算法求解模型参数,因此效率比GMM算法低,但与Gamma算法相比较,效率有很大的提高。

表 4 各算法的分割时间
Table 4 Segmentation time of each algorithm

下载CSV
GMMGammaGaMMHWGaMM
时间/s36.06691.74511.3490.72

2.2 实测SAR图像分割

图 5为实测SAR图像及各算法的分割结果。其中,图 5(a)为Radarsat-I卫星的海冰图像,其分辨率为8 m,采用C波段HH极化,尺寸为128×128像素。由暗到亮区域分别为海水,冰水混合物和海冰。图 5(b)中GMM算法难以将各区域分割开,尤其是较暗区域(冰水混合物)和亮区域中存在不同程度的误分割像素。图 5(c)中Gamma算法对亮区域内噪声较为敏感,其分割结果中存在不同程度的误分割。图 5(d)中GaMM算法的分割结果优于其他对比算法,但仍存在误分割区域,如图 5(d)的第3幅所示难以分割出图像中暗区域。图 5(e)中提出算法可将各区域分割开,各区域仅存在极少的误分割像素。综上,对实测SAR图像的分割,提出算法可得到最优的分割结果。

图 5 实测SAR图像及其分割结果
Fig. 5 Real SAR images and their segmentation results ((a)real SAR images; (b)GMM; (c)Gamma; (d)GaMM; (e)HWGaMM)

3 结论

为了实现高精度SAR图像分割,提出了一种空间约束HWGaMM的SAR图像分割算法。首先,HWGaMM的混合组份由多个Gamma分布加权和定义,以建模同质区域内像素强度复杂统计分布,且考虑到同质区域内像素强度的差异性和异质区域内像素强度的相似性,采用混合组份加权和定义HWGaMM。其次,为了提高分割算法的抗噪性,采用MRF建模像素邻域关系,利用中心像素及其邻域像素的后验概率定义混合权重。最后,提出算法结合M-H和EM算法求解模型参数,以提高分割效率。为了验证提出算法,对仿真和实测SAR图像进行分割实验。

从实验结果可得出以下结论:1)提出的HWGaMM具有准确建模复杂统计分布的能力;2)结合EM和M-H算法求解模型参数可提高分割效率;3)提出空间约束HWGaMM算法能够得到高精度的分割结果。分割算法的自适应性是学者们一直关注和不断研究的问题,为实现参数自适应性将进一步研究自适应HWGaMM分割算法。

参考文献

  • Gao G. 2010. Statistical modeling of SAR images:a survey. Sensors, 10(1): 775-795 [DOI:10.3390/s100100775]
  • Ji J, Lyu X J, Han L Y, Zhang C H. 2016. Fast and adaptive region merging based on perceptual hashing via multi-thresholding for SAR image segmentation. Remote Sensing Letters, 7(12): 1199-1208 [DOI:10.1080/2150704X.2016.1222097]
  • Ji J, Wang K L. 2014. A robust nonlocal fuzzy clustering algorithm with between-cluster separation measure for SAR image segmentation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7(12): 4929-4936 [DOI:10.1109/JSTARS.2014.2308531]
  • Shi X, Li Y, Zhao Q H. 2018. High resolution SAR image segmentation method based on hierarchical Gamma mixture model. Pattern Recognition and Artificial Intelligence, 31(7): 591-601 (石雪, 李玉, 赵泉华. 2018. 基于层次Gamma混合模型的高分辨率SAR影像分割方法. 模式识别与人工智能, 31(7): 591-601) [DOI:10.16451/j.cnki.issn1003-6059.201807002]
  • Song W Y, Li M, Zhang P, Wu Y, Jia L, An L. 2017. Unsupervised PolSAR image classification and segmentation using Dirichlet process mixture model and Markov random fields with similarity measure. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 10(8): 3556-3568 [DOI:10.1109/JSTARS.2017.2684301]
  • Tang D K, Wang F, Wang H Q. 2019. Single-polarization SAR data flood water detection method based on Markov segmentation. Journal of Electronics and Information Technology, 41(3): 619-625 (唐德可, 王峰, 王宏琦. 2019. 基于马尔可夫分割的单极化SAR数据洪涝水体检测方法. 电子与信息学报, 41(3): 619-625) [DOI:10.11999/JEIT180420]
  • Wang H M, Shi J Y. 2018. SAR image segmentation algorithm based on contourlet domain AFMRF model. IET Image Processing, 12(7): 1124-1130 [DOI:10.1049/iet-ipr.2017.0290]
  • Wang Y, Li Y, Zhao Q H. 2015. Segmentation of high-resolution SAR image with unknown number of classes based on regular tessellation and RJMCMC algorithm. International Journal of Remote Sensing, 36(5): 1290-1306 [DOI:10.1080/01431161.2015.1009655]
  • Wang Y, Li Y, Zhao Q H. 2016. SAR image segmentation combined regular tessellation and M-H algorithm. Geomatics and Information Science of Wuhan University, 41(11): 1491-1497 (王玉, 李玉, 赵泉华. 2016. 结合规则划分和M-H算法的SAR图像分割. 武汉大学学报(信息科学版), 41(11): 1491-1497) [DOI:10.13203/j.whugis20140536]
  • Xia M Q, Yang X Z, Dong Z Y, Zheng X, Li G Q. 2015. Segmentation of polarimetric SAR sea ice image with noise suppression. Journal of Remote Sensing, 19(5): 856-863 (夏梦琴, 杨学志, 董张玉, 郑鑫, 李国强. 2015. 噪声抑制的多极化SAR海冰图像分割. 遥感学报, 19(5): 856-863) [DOI:10.11834/jrs.20155028]
  • Xu L L, Wong A, Clausi D A. 2017. An enhanced probabilistic posterior sampling approach for synthesizing SAR imagery with sea ice and oil spills. IEEE Geoscience and Remote Sensing Letters, 14(2): 188-192 [DOI:10.1109/lgrs.2016.2633572]
  • Yu F J, Sun W Z, Li J J, Zhao Y, Zhang Y M, Chen G. 2017. An improved Otsu method for oil spill detection from SAR images. Oceanologia, 59(3): 311-317 [DOI:10.1016/j.oceano.2017.03.005]
  • Zhang H, Wu Q M J, Nguyen T M, Sun X M. 2014. Synthetic aperture radar image segmentation by modified student's t-mixture model. IEEE Transactions on Geoscience and Remote Sensing, 52(7): 4391-4403 [DOI:10.1109/TGRS.2013.2281854]
  • Zhang H T, Li Y N. 2015. Watershed algorithm with threshold mark for color image segmentation. Journal of Image and Graphics, 20(12): 1602-1611 (张海涛, 李雅男. 2015. 阈值标记的分水岭彩色图像分割. 中国图象图形学报, 20(12): 1602-1611) [DOI:10.11834/jig.20151205]
  • Zhang J J, Li Y, Zhao Q H. 2016. Combining the expection maximization and genetic algorithms for SAR image segmentation within a MAS framework. Journal of Image and Graphics, 21(1): 86-94 (张金静, 李玉, 赵泉华. 2016. 多主体框架下结合最大期望值和遗传算法的SAR图像分割. 中国图象图形学报, 21(1): 86-94) [DOI:10.11834/jig.20160111]
  • Zhang Y H, Li Y, Zhao X M, Zhao Q H. 2016. ECM-based segmentation for multi-look SAR image. Journal of Remote Sensing, 20(6): 1391-1401 (张英海, 李玉, 赵雪梅, 赵泉华. 2016. ECM算法的多视SAR影像分割. 遥感学报, 20(6): 1391-1401) [DOI:10.11834/jrs.20165288]
  • Zhao Q H, Li X L, Li Y. 2017. Multilook SAR image segmentation with an unknown number of clusters using a Gamma mixture model and hierarchical clustering. Sensors, 17(5): 1114 [DOI:10.3390/s17051114]
  • Zhao Q H, Shi X, Wang Y, Li Y. 2017. Remote sensing image segmentation based on spatially constrained Gaussian mixture model with unknown class number. Journal on Communications, 38(2): 34-43 (赵泉华, 石雪, 王玉, 李玉. 2017. 可变类空间约束高斯混合模型遥感图像分割. 通信学报, 38(2): 34-43) [DOI:10.11959/j.issn.1000-436x.2017026]
  • Zhao X M, Li Y, Zhao Q H. 2014. Image segmentation by fuzzy clustering algorithm combining hidden Markov random field and Gaussian regression model. Journal of Electronics and Information Techno-logy, 36(11): 2730-2736 (赵雪梅, 李玉, 赵泉华. 2014. 结合高斯回归模型和隐马尔可夫随机场的模糊聚类图像分割. 电子与信息学报, 36(11): 2730-2736) [DOI:10.3724/SP.J.1146.2013.01751]
  • Zhen H, Guo W, Yan J W. 2019. Synthetic aperture radar image target segmentation method based on boundary and texture information. Journal of Image and Graphics, 24(6): 882-889 (谌华, 郭伟, 闫敬文. 2019. 综合边界和纹理信息的合成孔径雷达图像目标分割. 中国图象图形学报, 24(6): 882-889) [DOI:10.11834/jig.180484]