Robust non-local fuzzy c-means algorithm with edge preservation for SAR image segmentation

Fuzzy c-means (FCM) algorithm has been proven effective for image segmentation; nevertheless it is sensitive to different types of noises. Up to now, a series of improved FCM algorithms incorporating spatial information have been developed, which are robust for Gaussian, uniform, and salt and pepper noises. However, limited effort has been placed on tackling the problem of a large amount of intrinsic and undesired multiplicative speckle in synthetic aperture radar (SAR) images. A crucial problem for SAR image segmentation is to guarantee speckle insensitiveness and edge detail preservation simultaneously. To address this problem, a robust and specific non-local FCM algorithm with edge preservation for SAR image segmentation is proposed. In this study, a new image is constructed using the non-local information and rectifying the edge parts, which is robust for speckle without sacrificing edge sharpness. To measure the patch-similarity in non-local method effectively, a novel generalized ratio distance based on SAR multiplicative speckle is defined. To locate and rectify the edge parts, coefficient of variation (CV) based threshold and orientation based statistics methods are designed. At last, this new image is clustered by FCM algorithm. Compared with six improved FCM algorithms and two state-of-the-art segmentation algorithms (spectral clustering and normalized cuts), the proposed algorithm obtains the best performance in terms of region uniformity and boundary localization.

► Non-local information is used to construct a new speckle-insensitive image. ► Edge part of the new image is rectified to preserve its geometric structure. ► NLEP-FCM performs well in region uniformity and boundary localization. ► NLEP-FCM reduces the time complexity and avoids some parameters.