基于改进EM算法的高斯混合模型图像聚类方法

Gaussian Mixture Model Image Clustering Method Based on Improved EM Algorithm

针对现有高斯混合模型对初始值敏感并容易陷入局部最优值的情况,提出一种基于改进EM算法的高斯混合模型图像聚类方法。该方法首先使用惯性权重先增后减的粒子群完成高斯混合模型参数初始化;然后引入近似骨架理论对EM算法进行改进优化,求解出高斯混合模型的最终参数;最后在图像聚类应用中进行仿真实验。基于改进EM算法的高斯混合模型图像聚类方法,结构相似度相较于标准EM算法、RSEM算法和PSOEM算法分别提升6.31%、4.20%和1.38%;余弦距离值相较于标准EM算法、RSEM算法和PSOEM算法分别提升4.12%、2.69%和0.94%。实验结果表明,该方法能够有效提升局部像素区域的聚类效果,获得聚类边界更加清晰的输出图像。

Aiming at the situation that Gaussian mixture model is sensitive to the initial value and easy to fall into the local optimal value, a Gaussian mixture model image clustering method based on the improved EM algorithm is proposed. The method first uses the particle swarm with the inertial weight to increase and then decrease to initialize the parameters of the Gaussian mixture model, then introduces the approximate skeleton theory to improve and optimize the EM algorithm, solve the parameters of the Gaussian mixture model, and finally perform simulation experiments in the image clustering application. SSIM values of gaussian mixture model image clustering method based on improved EM algorithm are 6.31%, 4.20% and 1.38% higher than those of standard EM, RSEM and PSOEM, respectively. The cosine distance is 4.12%, 2.69% and 0.94% higher than that of which. Analysis of experimental results shows that the method can effectively improve the clustering effect of local pixel regions and obtain a clearer output image with cluster boundaries.