Abstract: Finite mixture model under Laplacian distribution is proposed in order to solve the problem that image segmentation based on Gaussian Mixture Model（GMM） failing to settle tailing situation with heavy-tailed noise, besides, unlike the standard Laplacian Mixture model（LMM） where pixels themselves are considered independent of each other, the proposed method incorporates the spatial neighborhood relationship of pixels into the standard LMM. In order to estimate model parameters from observations and instead of utilizing an expectation-maximization algorithm, the gradient method is adopted. The experimental results demonstrate the robustness, accuracy, and effectiveness of the method in comparison with the standard LMM and GMM.

Key words: Laplacian Mixture Model（LMM）, image segmentation, heavy-tailed noise, spatial neighborhood relationship

摘要： 针对高斯混合模型（GMM）不能有效处理重尾噪声下图像拖尾情况，提出了基于拉普拉斯（Laplacian）分布的有限混合模型图像分割方法。与标准拉普拉斯混合模型（LMM）将像素点作为孤立个体不同的是，该方法充分考虑了相邻像素点间的空间关系。相较传统混合模型参数估计采用的EM算法，该方法采用梯度下降法优化参数。实验结果表明在处理重尾噪声时，该方法与标准LMM算法和GMM算法相比，鲁棒性更好，分割更精确有效。

关键词: 拉普拉斯混合模型（LMM）, 图像分割, 重尾噪声, 空间邻域关系