Speckle Suppression Of SAR Images And Extraction Of Local Features

Since 90's of the last century, SAR application has made a great progress, and several SAR systems, for example, European ERS, Canadian Radarsat, etc., were launched successively. These systems have been observing the ground for a long time. In our country, various SAR equipments have been in applications and by thesee equipments huge SAR data are obtained. It is an important issue how to utilize these radar data and SAR images to serve for military requirements and national economy construction effectively and efficiently. With maturity of SAR imaging technique, the information processing technique of SAR images are widely investigating. Improving SAR image quality using signal processing methods and extracting the geometric features in SAR images are two indispensable pre-processing steps for SAR image information processing.In this dissertation, we investigate speckle suppression method in SAR images using wavelet transforms and local geometric feature extraction algorithms in SAR images. The dissertation mainly includes the four pasts as follows:(1) SAR imaging using refined matched filters. In SAR imaging, the matched filters are often used to realize the high resolution in range and azimuth dimensions. It is known that the matched filters are based upon the maximal output signal-to-noise ratio (SNR) criterion. SAR image understanding and information processing require that the imaging algorithm can preserve true details of scenes and avoid'false'details except for high resolutions and input SNR. Obviously, the matched filter is easy to generate some'false'details owing to its high sidelobe and sharp vibration of the matched filtering output. Referring to Canny's edge detection idea in image processing, the Canny criteria, (1) maximal output SNR, (2) location constraint, and (3) less oscillation, are used to improve the matching filters in SAR imaging. The obtained refined matched filters can effectively reduce the'false'details around strong scatterers.(2) SAR image despeckling via wavelets. Speckle noise in SAR images is the main factor to degrade the quality of SAR images and exhibits multiplicative noise. As a result, the matured wavelet-based image denoising methods for additive noise in common optical images cannot be directly applied to suppress speckle noise in SAR images. Here, we use the spatial inhomogenous additive noise to model speckle noise in SAR images. Several wavelet-based despeckling methods are proposed. First, a simple Bayssian shrinkage method in wavelet domain is developed, where spatial-varying noise powers in wavelet domain are estimated by the variance distribution in pixel domain and wavelet coefficients of SAR images are modeled by the generalized Gaussian distributions. Second, the doubly Wiener filtering algorithm in wavelet domain are extended to improve the performance of speckle noise suppression in SAR images. Third, considering the non-symmetry of wavelet coefficients of SAR images in amplitude, we propose the wavelet-based despeckling method based on non-symmetric generalized Gaussian distributions. Compared with the traditional spatial domain filtering methods and the existing wavelet-based methods, the proposed methods achieve better overall performance.(3) SAR image despeckling via nonlinear filtering. In the wavelet-based methods for SAR image despeckling, detail preservation and smoothness in homogenous regions in SAR images are always conflictive with each other. In SAR images, strong speckle pixels exhibit impulse-noise characteristics and thus these noise pixels are difficult to be removed by the linear smoothing filtering. It is known that various median filtering methods are suited for removing impulse noise in images and can better preserve details in images such as edges. Here, we propose the median filtering algorithm based on local geometric structures to suppress speckle noise in SAR images. The new algorithm combines the traditional spatial-domain Gamma MAP algorithm and the detail-preserved median filtering algorithm. A SAR image is first despeckled with the Gamma MAP algorithm. The filtered image contains a few'bright'and'dim'pixels different from most pixels around it. These pixels are probably residual impulse noise or belong to the edges of the SAR image. Here, we use the local geometric structure to classify these pixels into impulse noise and detail pixels. The median filtering is applied to the impulse noise to remove it while detail pixels remain invariant. The experimental results show that the proposed nonlinear filtering method better preserves details in SAR images when speckle noise is suppressed.(4) Accurate edge localization and local oriented information extraction in SAR images. Edge localization and local oriented information extraction in SAR images is crucial in searching objects of interest from complicated SAR images. It has been shown that the traditional accurate edge localization based on rectangle subwindows gives a biased estimate and estimated edge always leans to the dark part in SAR images. Here, we give the more accurate edge localization method based on sandglass-shape window. Additionally, the discrete steerable filters based on wavelets are proposed to extract the local oriented structure of a pixel in images. This local oriented structure is useful in point-like target classification.