Segmentation Of SAR Image Based On Mixture Multiscale ARMA Model

Synthetic aperture radar (SAR) image segmentation is a key technique of automatic target recognition (ATR) and SAR image information processing. The presence of speckle on SAR images not only reduces the interpreter's ability to resolve fine detail, but also makes automatic segmentation of such images difficult, either by gray levels or texture. In this paper, we present several new segmentation methods based on multiscale autoregressive moving average (ARMA) model and mixture multiscale ARMA model, which can fully exploit the variations in speckle pattern as image resolution is varied from course to fine. The details are as follows:Firstly, we investigated the stationary conditions of multiscale ARMA model, and proposed two unsupervised segmentation methods based on MAR and multiscale ARMA model. These models capture the statistical information in a multiscale sequence of SAR image, which is then used to implemented segmentation of SAR image via multiresolution mixture algorithm. In order to evaluate the performances of our methods, we analyzed the capability of MAR model and multiscale ARMA model in extracting information from SAR images.Secondly, we further combined the information extracted based on multiscale ARMA model with multiscale Markov random field (MRF) model. MPM algorithm is used to segment the SAR image. Experimental Results show that the proposed method not only overcomes the limitation of multiscale MRF model, but also preserves the low computational complexity.Finally, we generalized the multiscale ARMA model to mixture multiscale ARMA model, which is a natural mixture multiscale framework for the modeling of non-linear statistical signal and image. The stationary conditions are derived. In addition, a mixture multiscale ARMA network is proposed for unsupervised segmentation of SAR image. The network combines the multiscale analysis (MA)method and the feedforward artificial neural network (FANN), thus it maintains some of the characteristics of the MA method and the FANN respectively. A corresponding learning algorithm is derived based on the Bayes information criterion (BIC) and genetic algorithm (GA). Experimental results on SAR images are shown to validate the presented network and learning algorithm.