Research On Features Extraction And Targets Recognition In SAR Images

SAR is of important use for military reconnaissance and civil activity, so it's meaningful and has application prospect to study feature extraction and object recognition method of SAR images. In this dissertation, we detailedly analyze the feature extraction and object identification methods of SAR images in theory, and validate them by experiment.We study multiscale edge detection of SAR image. According to the transfer characteristic across scales of the wavelet module of signal edge and noise edge, we combine the property of edges in different scale and propose multiscale edge fusion algorithm consisting of edge transfer, edge inherit and edge growth. The result of experiment shows that this algorithm can get ride of the affect of noise and the edges fused have precise position and intact contour. We realize automatic extraction of road network in SAR images by a two-step algorithm. We first propose an improved local edge detector to detect road elements from SAR images. By introducing prior information of road network, we can organize the local line fragment into useful line structure with directional potential function. The algorithm is applied to road network detection in ERS-1 image, and good result is obtained.We study SAR image filtering based on wavelet domain hidden Markov model (HMM). A wavelet domain hidden Markov tree (HMT) model is constructed to model statistical dependence and nonGaussian statistics of wavelet coefficient. The estimate of HMT model parameters can be obtained by EM algorithm. After converting multiple speckle noise to additional Gaussian noise, we achieve the MMSE estimate of SAR image wavelet coefficient. We propose a kind of SAR image filtering method based on iterated structure filter. This method introduces a series of neighborhood models which reflect local edge property to describe image details. The prior information of pixel intense distribution is introduced. Then simulated annealing algorithm is applied to choose the proper neighborhood structure, and the optimal estimate can be obtained. Both filtering methods can significantly reduce speckle noise, and at the same time preserve the image notable details.We propose a SAR image segmentation method based on the criterion of likelihood difference. This method utilizes the distribution property of image intensity. Given the false alarm probability, the threshold for likelihood difference can be determined. Merging cost which reflect region structure property is defined to control merging order. In order to reduce calculation burden, and at the same time keep image details, we propose optimal initial segmentation method. We construct cost function which combines the likelihood function and boundary constraint function. On the basis of block initial segmentation, Simulated annealing algorithm is used to get the optimal initial segmentation which minimizes the cost function.We propose a multistage SAR target recognition process based on target detection, target segmentation, target aspect estimation and invariant feature extraction. First, we propose a kind of constant false alarm rate (CFAR) target detector to find potential targets from scenario. Considering the relationship of shadow, background and target, we propose proper MRF model to describe image prior information. The Bayes formulation is adopted to get the MAP segmentation. Sigularity values of target contour and target signal are used to represent the invariant features of target. A feature template set is constructed for every target class, and every feature template is identified by its aspect. We propose a synthetical method to estimate target aspect from segmented target region. The feature templates whose aspects are close to the estimated aspect are selected. Feature Matching criterion based on sigularity values of target contour and target signal is given to identify target. Experimental results show that this method can achieve high recognition probability.