Unsupervised SAR Image Segmentation Based On Region Mergin

Synthetic Aperture Radar (SAR) is a high-resolution coherent imaging radar andeffective to identify the camouflages and penetrate the covers, operating day and nightunder any weather conditions. It breaks through the limitation of optical remote sensingby weather and other external conditions affecting.Unlike optical images which typically include multiple bands of gray informationfor easy to target recognition and classification, SAR images only record one band ofecho information with binary plural form, which can be converted to relevant amplitudeand phase information. The amplitude information usually corresponds with backscatterintensity of ground targets for radar wave and closely related to target media, moistureand roughness, which has a strong correlation with gray information of optical images.Because of the relatively low image resolution and high noise ratio, the amplitudeinformation in SAR image is far less than that in optical image of same level. At thesame time, SAR images are corrupted with multiplicative speckle noise, which is primarily dueto the interference of coherent waves, and geometric distortion. These degrade fine details andedges of the objects present in the scene, which leads to difficult image processing tasks.This paper proposes two unsupervised SAR image segmentation algorithm based on regionmerging, and the backscatter information are combined with edge information and textureinformation to obtain a better image segmentation result. The algorithms are presented below.(1) Proposing an unsupervised SAR image segmentation based on superpixelsmerging in ICA independent space (ISRMS). The final number of clustering is the onlyparameter in the new method, and no prior information is necessary. Firstly, anindependent space is proposed to represent SAR images, which is obtained byIndependent Component Analysis. It maps the SAR image into three independentspaces for extending the single channel to three channels and enriching the informationof SAR image. Secondly, the image is divided into small regions by superpixelsalgorithm, then the feature vectors of small regions are extracted in the independentspace for similarity calculation. Thirdly, those small regions are merged in regionadjacent graph (RAG) and Full-connected graph (FCG) based on the Mining SpanningTree theory, which balances the speed and quality of segmentation. Finally, theperformance of ISRMS is proved in the simulated image and real SAR imagescompared with Graph-cut and SLIC. Moreover, the effectiveness of each stage of thealgorithm is also validated through experimentation. (2) Proposing an adaptive SAR image segmentation with histogram-basedperceptual hashing (HpHS). It is an unsupervised algorithm which can automaticallyobtain the final number of clusters, not artificially set. In the preprocessing stage ofgeneral segmentation algorithm based on region merging, the SAR image is divided intoconnected regions with closed boundaries, while the HpHS uses the histogram for raydegradation, and regard each gray level as one original region, for reducing theinfluence on results of initial boundary. Then, the perceptual hashing, which is appliedin matching thumbnail, is introduced into image segmentation in order to calculate thesimilarity between regions with structure information. Finally, four group experimentsare designed to analyze the performance of HpHS algorithm, and the results show thatthe HpHS algorithm’s error rate is less than15%.