基于复小波域树结构化MRF模型的声纳图像分割

声纳图像受成像环境影响对比度低,特性信息较弱,且图像分辨率不高,用传统的分割方法效果较差,为此,构建了双树复小波域树结构化的MRF模型(TS-MRF),提出了基于此模型的声纳图像分割算法。双树复小波变换(DT-CWT)具有近似平移不变性和良好的方向选择性,其多分辨率分析能有效地提取声纳图像的弱特征信息,以便TS-MRF中节点参数的描述能准确地反映树结构的分布规律和图像统计特性;复小波域6个方向高频子带相互独立,尺度间父子节点标号具有一阶Markov性;尺度内构建TS-MRF模型,且每一节点标号依赖于父节点,采用Potts模型对节点标号势函数建模,相同标号的观测特征用高斯模型建模;最后,用多分辨率递归和每一分辨率分类层次树从顶层向底层的尺度内递归算法来求解最大后验概率,实现分类层次树标号,完成声纳图像分割。实验结果从视觉效果和定量分析两方面验证表明,本文算法能有效地抑制噪声,较好地提取声纳图像的弱特征信息,具有较高的分割精度和鲁棒性。

Sonar image segmentation based on tree structured MRF model in complex wavelet domain

Due to the impact of the imaging environment, the sonar images have the characteristics of low contrast, weak feature information and low image resolution, which leads to poor segmentation effect for traditional segmentation methods. Aiming at this problem, the tree structured MRF (TS MRF) model in dual tree complex wavelet domain is constructed, a new sonar image segmentation algorithm is proposed based on this model. Dual tree complex wavelet transform (DT CWT) has approximate shift invariance and nice directional selectivity, with multi resolution analysis it can effectively extract the weak information characteristics of sonar image. So that the node parameters in the TS MRF model can accurately reflect the tree structure distribution and image statistical properties: the high frequency sub bands in six directions in complex wavelet domain are mutual independent, the labels of the father child nodes in the inter scale have the property of first order Markov. The TS MRF model is constructed in the intra scale, and for each node the label depends on its parent node; the Potts model is adopted to build the model for the potential function of the node label; and Gaussian model is used to build the model for the observed characteristics with the same label. Finally, the intra scale recursive algorithm is used to get the result of the maximum a posteriori probability from the top to the bottom of the multi resolution recursion and each resolution classification hierarchy tree, to achieve the classification hierarchy tree labeling and complete the sonar image segmentation. From two aspects of visual effect and quantitative analysis, the experiment results indicate that the algorithm not only can effectively suppress noise but also can better extract weak characteristic information of sonar image, and has higher segmentation accuracy and robustness.