融合局部和非局部信息的自适应贝叶斯分割方法

摘要

With only considering the impact of neighborhood pixels, the traditional Bayesian segmentation method based on Markov Random Field (MRF) can not suppress the speckle noise effectively. In the traditional priori model, the influence of each pixel within the neighborhood to the center one is assumed the same, which makes the description of the edge imprecise and the segmentation ineffective. Thus, an adaptive Bayesian segmentation method fused of local and non-local information is proposed. For the multiplicative noise model contained in SAR image, the similarity measure based on ratio probability is introduced, and the nonlocal similar pixel-blocks are adopted to guide the segmentation of the current pixel. Furthermore, the Coefficient of Variation (CV) method is employed to obtain the image template of edge area. In the edge region, the structure index and the size of search window are adaptively adjusted to improve the inconsistency between excessive smooth and structure preserving. In the experimental analysis, parts of the SAR image segmentation results with the new technique are given, which are compared with the traditional means. There is a significant advantage that the proposed algorithm enables more accurate segmentation results, which not only make the speckle noise suppressed, but also keep the detail characteristics effectively.%传统基于马尔可夫随机场(MRF)的贝叶斯分割方法由于只考虑邻域像素点的先验影响，无法有效抑制相干斑噪声；边缘区域分割效果欠佳，因为先验模型假定邻域中每个像素对中心像素的影响相同。因而，该文提出一种融合局部和非局部信息的自适应贝叶斯分割方法。针对SAR图像中的相干斑噪声模型，引入基于比率概率的相似性测度，用非局部相似像素块指导当前像素点的分割；并且采用变分系数(Coefficient of Variation, CV)方法获取边缘区域图像模板，在边缘区域自适应地调整定义的结构指数以及搜索窗尺寸，从而改善分割过度平滑与结构保持的矛盾；在实验分析中，利用新方法对部分图像进行了分割实验，并与传统方法作了比较。改进方法的分割结果形状更为准确，不但抑制了相干斑噪声，还有效保持了细节特征，具有显著优势。