Research On SAR Image Segmentation Algorithm Based On Texture Feature And Superpixels

Synthetic Aperture Radar(SAR) is a system with all-weather,high-resolution,large-width,unconstrained by climatic conditions,and surface penetrability,which makes SAR have unique advantage in disaster monitoring,environmental monitoring,marine monitoring,resource exploration,crop production estimation,surveying and military.As an important research topic of image interpretation systems,the goal of image segmentation is to segment the ground objects in the image as completely as possible to provide support for subsequent classification and application.Although the introduction of superpixels thought has brought many effective methods and has made great progress in image segmentation,there are still many problems in the field of SAR image segmentation.The existing mature superpixels algorithms that are all designed for optical images and sensitive to the noise in SAR images,so they cannot work well if just direct use them in SAR images.In addition,in the existing segmentation algorithms,single image feature cannot segment the entire content of the image,but multi-feature segmentation would bring the problem of high computational complexity,which cause some contradictions.Therefore,in order to improve the speed of segmentation while segmenting the image better,two efficient SAR image segmentation algorithms are proposed in this paper: SAR image region merging segmentation method based on improved LGRP and superpixels,and SAR image clustering segmentation method based on texture features and edge constraints.The two algorithms are as follows:(1)A SAR image segmentation algorithm that based on improved LGRP and superpixels is proposed.Considering the problem that the existing superpixels algorithms easily affected by the multiplicative noise of SAR images,this segmentation algorithm introduces a SAR image texture feature with good noise robustness: local gradient ratio mode feature,to overcome the effect of noise and get a better segmentation result.At the same time,because the texture feature and gray feature are a set of mutually beneficial local feature expression combination,this algorithm integrates these two local features and introduces them into a simple non-iterative clustering superpixels generation algorithm(SNIC)while improving its similarity metrics,to generate superpixels that are more suitable for SAR images.Then introduces the initial result of superpixels into the region merging process,and uses the nearest neighbor graph and the improved heterogeneity criterion to merge them and obtain the segmentation result.Finally,validating the method in this paper through multiple sets of comparative experiments.(2)A SAR image clustering segmentation algorithm that based on texture feature and edge constraints is proposed.Though the superpixels algorithms can greatly improve the segmentation's speed,it is difficult to perfect segment images just based on texture feature and grayscale feature alone.Based on the previous algorithm,this paper proposes a superpixels generation algorithm based on multi-neighborhoods texture feature.It uses the Gaussian gamma double window detector combined with morphological reconstruction to detect image boundaries.The role of edge information is to guide the generation of initial superpixels.Then we extract the local texture and grey feature histograms of superpixels and use fast fuzzy clustering segmentation algorithm for clustering.Experiments show that introduces edge information can enhance the ability to detect the edge of the target,and at the same time can obtain more accurate segmentation result.