基于Contourlet域HMT模型的声纳图像去噪

声纳图像预处理是声纳图像目标识别与跟踪的前提;声纳图像对比度低,特性信息弱,为此,提出Contourlet域HMT模型(CT-HMT)的声纳图像去噪算法。Contourlet域中,不同方向间子带系数的相关性体现于DFB分解中,相邻尺度间父节点对应的4个子节点分布在2个可分离的方向子带上,父、子节点状态"持续性"采用Markov模型建模,尺度内Contourlet系数的"聚集性"采用混合高斯模型建模;最后,用贝叶斯准则估计无噪图像的Contourlet系数,实现声纳图像去噪。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,提取声纳图像的弱特征信息,较好地保全了图像的边缘和轮廓信息。

Sonar Image De-noising Based on HMT Model in Contourlet Domain

Sonar image preprocessing is the precondition of object recognition and tracking. For the impact of imagery environmental factor, the sonar image has the disadvantages of low contrast ratio, weak feature information, the coarse resolution and so on. Compared with common optical image, it is difficult to get well de-noising results by using traditional de-noising algorithm. Therefore, a sonar image de-noising algorithm based on Contourlet domain HMT (CT-HMT) model is proposed. Contourlet analysis not only has inherited the multiresolution analysis characteristic of wavelet, but also has better directional selectivity, which can effectively extract the weak characteristic information of sonar image. It is a foundation for the parameters of the parent-child nodes to accurately reflect the distribution of tree structure and statistical property of sonar image in the HMT model. In Contourlet domain, the correlation of the sub-band coefficients between different directions is embodied in the DFB decomposition. Between adjacent scales, the four corresponding child nodes of the parent node are distributed on two separable sub-bands, and the status of parent-child has persistent property of first-order Markov. The aggregation of the Contourlet coefficients on the intra-scale is modeled by mixture gauss model. Finally, estimate the Contourlet coefficients of the no-noise image depending on Bayesian principle to realize the sonar image de-noising. From two aspects of visual effects and quantitative analysis, the experimental results show that the algorithm can effectively suppress noise and extract weak information of sonar image, and can better keep the edge and contour information of the image.