基于K-GMM算法的SAR海冰图像分类

SAR Sea Ice Image Classification Based on the K-GMM Algorithm

为适应SAR海冰图像的斑点噪声和入射角效应,已有的基于随机场模型的海冰分割算法常引入复杂的语境模型,容易导致过平滑现象,不仅降低了部分图像的分割精度,而且运行效率较低。该文依据海冰业务化分类(分割)的特点,提出了一种基于GMM与K-均值的改进聚类方法:首先,在K-均值聚类步骤中集成了合并操作,为基于GMM的聚类提供初值,而且大大缩减了GMM聚类步骤的迭代次数,提高了输出结果的稳定性;然后,利用EM方法学习GMM,完成ML分类,将每个像素分配到最终类别中。实验结果表明,在分类精度可比甚至更高的情况下,与其他几种常见算法相比,该方法有效降低了分类算法的计算复杂度,减少了对计算资源的需求,易于移植到小型化、移动化设备上。

Due to the speckle and incident angle effect of the SAR sea ice images,the existing sea ice segmentation algorithm based on the random field method often introduces the complex context model,which can easily lead to the over smoothing phenomenon.These methods not only reduce the segmentation accuracy of some images,but also compromise the program operational efficiency.Therefore,according to the characteristics of sea ice classification(segmentation),this paper proposes an improved clustering method based on Gaussian mixture model(GMM)and K-means clustering.The merging operation is integrated in the K-means clustering step,which provides initial values for the clustering based on Gaussian mixture model.On this basis,the GMM is learned using the expectation maximum(EM)method.And finally,an maximum likelihood(ML)classification is performed to assign each pixel into a final class.Among them,the introduction of the K-means clustering step greatly reduces the number of iterations of the GMM clustering step and improves the stability of the output result.The proposed method was tested on various SAR images,and the results were compared with other well-established approaches.The quantitative analysis of the experimental results confirmed that the proposed method can significantly lower the computational complexity of the classification algorithm,decrease the amount of computing resources on demand when the classification accuracy is comparable or even higher.Also,the proposed method is easy to implement in miniaturized or mobile devices.