Superpixel Segmentation Algorithms For SAR Images Based On SLIC

Synthetic Aperture Radar(SAR)is an important device for remote sensing of the earth.It can measure the surface of the earth all day and all-weather,and then extracts the physical information and parameters of the imaging surface.Polarimetric SAR is a kind of SAR that can work in a variety of polarized combinations at the same time.It has the capability of acquiring more information than conventional single polarized SAR,and contains more information of the ground objects.Therefore,SAR image interpretation is of significance.Generally speaking,methods of SAR image interpretation can be divided into two categories from the perspective of processing basic unit,including: 1)pixel-based methods,2)region-based methods.The traditional pixel-based algorithm is sensitive to speckle noise and is of high computational complexity.The region-based algorithm overcomes these shortcomings and makes full use of the information of region.Therefore,the segmentation of region in SAR images plays an important role in the subsequent processing and interpretation of SAR images.Superpixel segmentation is an image segmentation algorithm,which is a hotspot in the current image segmentation.The superpixel refers to a uniform and compact image area with an irregular shape,which is obtained after an image is over-segmented.As the superpixel segmentation can effectively reduce the image complexity and shorten the processing time,the superpixel segmentation is often used as an pre-processing step of image interpretation.Many superpixel segmentation algorithms are designed for optical images.Because of the existence of coherent speckle noise in SAR images,these algorithms can not be directly applied to the superpixel segmentation of SAR images.Therefore,this paper focuses on SAR image superpixel segmentation based on Simple Linear Iterative Clustering(SLIC)algorithm.The main research work and achievements are shown as follows:1)A new method of single-polarization SAR image superpixel segmentation based on SLIC and generalized Gamma distribution is proposed.SLIC algorithm is a very popular optical image superpixel segmentation algorithm proposed in recent years.However,when the algorithm is applied directly to SAR images,the generated superpixels are not good due to strong speckle noise and large dynamic range of pixel gray values in SAR images.To solve these problems,this paper proposes a superpixel segmentation algorithm for SAR images that can be applied to SAR images.This algorithm combines the spatial Euclidean distance and likelihood information estimated by the generalized Gamma distribution as the similarity measure between the pixels for local clustering.In the post-processing,the local edge evolving scheme combining the spatial context information and the likelihood information is adopted to improve the edge fitting degree of the generated superpixels.The experimental results show that the algorithm can generate superpixels of SAR images with high edge-fitting degree without considering the computational time complexity.2)A local iterative clustering algorithm for polarimetric SAR image superpixel segmentation is proposed.Because of the influence of speckle noise in polarimetric SAR images,SLIC can not be adopted directly.In order to solve this problem,this paper proposes an improved clustering center initialization procedure and an improved postprocessing step based on SLIC.In the local K-means clustering,the weighted sums of the revised Wishart distance and Euclidean distance are used as the distance measure between the pixels and clustering centers.The class label of the pixels is reassigned to the nearest cluster center in the local searching area.The segmentation result of the polarimetric SAR images in then processed by an improved post-processing step to merge the isolated small-sized superpixels into the larger superpixels which are the most similar to each other.The experimental results show that the proposed algorithm can generate uniform and superpixels for polarimetric SAR images with better edge preserving.3)This paper proposes a superpixel segmentation algorithm for polarimetric SAR images based on SLIC and log-Euclidean distance.Since each pixel in a polarimetric SAR image can be represented by a covariance matrix(or coherent matrix),and the covariance matrix can be considered as essentially a Riemannian manifold.Therefore,it can be measured using the distance measure of the Riemannian manifold Similarity in Polarimetric SAR Images.Based on this,this chapter exploits the log-Euclidean distance between covariance matrices and calculates the distance between the polarimetric SAR pixels by combining the spatial Euclidean distance and log-Euclidean distance to obtain superpixels.In addition,in order to obtain a uniform superpixel with a large size and a high edge preserving as well as retaining a strong point target,a new post-processing strategy is adopted in the post-processing stage to iteratively merge similar superpixels.A measure of dissimilarity is defined to preserve the strong target points in polarimetric SAR images.The experimental results show that the proposed algorithm can generate uniform superpixels of polarimetric SAR images and prserve the strong target points well.