基于均值漂移与空间信息的导向模糊C均值遥感图像分割算法

Guided fuzzy c-means clustering with mean shift and spatial information for remote-sensing image segmentation

摘要:

作为图像数据结构分割的重要工具,模糊C均值已被广泛应用于计算机视觉领域。然而模糊C均值在图像分割过程中不能有效地保留边缘和抑制噪声,往往得不到理想的分割结果。为解决这一问题,本文利用导向滤波器推导出一种新的改进模糊C均值算法。该算法的第一个创新点是其线性平移不变滤波过程,利用边缘保持平滑特性来保留分割中的边缘结构。第二个创新点是该技术通过将空间信息引入目标函数来改善对噪声的鲁棒性,空间信息通过导向滤波的平均输出获得。为了解决聚类算法中初始聚类中心问题,在图像分割过程中使用均值漂移算法选取初始聚类中心。本文方法的主要优点在于其对边缘保留和噪声具有鲁棒性,进而提高分割精度。基于合成图像和真实遥感图像的实验结果表明,与其他主流分割算法相比,该方法在分割性能方面表现出了良好的性能。

Abstract:

Fuzzy C-means (FCM) has widely been applied to computer vision, which emerged as an important tool for segmenting the structure of image data. However, the effectiveness of this technique lies in its inability to preserve edges and suppress noise, often leading to unsatisfactory segmentations. To solve this problem, we derive a modified FCM algorithm by using guided filter. The first key concept of our method is its linear translation-variant filtering process, which exploits edge-preserving smoothing property to preserve the edge structures in segmentation. The second is that this technique improves the robustness to noise by incorporating the spatial information into the objective function, which are obtained by the mean output of guided filtering. Third, mean shift algorithm is used to get initial cluster centers so that the algorithm does guarantee convergence to the global optimum. The main advantages of the proposed method are that it exhibits robustness to edge-preserving and noise and it can enhance the segmentation accuracy. By comparing with other segmentation methods, experimental results on both synthetic and real remote-sensing images suggest that the proposed method behaves well in segmentation performance.