Sketching Model And Steerable Kernel Function Based SAR Speckle Reduction

Due to the obtained at any-weather and any-time with high resolution, Synthetic Aperture Radar(SAR) imagery has been applied in many aspects, such as national defense, resource surveying, disaster monitoring and urban planning etc. However, owing to the coherence based imaging in SAR imaging system, SAR imagery is always accompanied with some stronger grain-like spots which are named as speckle. For the existence of stronger speckle in SAR imagery, some important targets are always very difficult to be discriminated. Thus, for the visualization and the auto-interpretation of SAR imagery, the problem of SAR speckle reduction was proposed at the beginning of the study on the SAR image processing. Moreover, speckle reduction has become one of important techniques in SAR image processing.In order to make an effective reduction of speckle and recovery the true value of SAR image, our thesis was unfolded from the geometric-structural property and the statistical property of SAR image. All works contained in the thesis are summarized as:(1) Inspired by the theory of vision computation theory proposed by Marr and the primal sketch model designed for natural images, a SAR sketching model and corresponding sketching method are proposed by considering the statistical property of SAR image. In the 1980 s, by summing up the results obtained on the human vision from psychophysics, neurophysiology and anatomy, Marr pointed out that human vision was essentially the information processing, and built a framework for the vision computation. As the earliest theory on compute vision, the vision computation theory is the solid foundation for latter technical development of the compute vision and image processing. Latterly, based on the primal sketch theory in Marr’s vision computation theory, Guo et al. proposed a primal sketch model and designed an effective method to extract the sketch graph of nature images. Moreover, the designed method was applied for the compressing and the reconstructing of natural images. In Chapter 2, inspired by the primal sketch method proposed by Guo et al., the statistical property of SAR image is considered and a SAR sketching model and corresponding sketching method is proposed to extract the sketch map of a SAR image to represent the geometric-structural features contained in the SAR image, such as edges and lines. The experiments on the SAR images of different resolutions, demonstrate that the obtained sketch map is effective on the representation of the geometrical properties(e.g. orientation and location) of the stick-like features in SAR images. Moreover, the proposed method is robust on the multiplicative speckle contained in SAR images.(2) Given the ignorance of geometrical property of the image patch in patch-based similarity, SAR sketching method proposed in Chapter 2 is introduced in Chapter 3 and a sketch based geometrical kernel function is designed for the local homogeneous region based SAR speckle reduction. In this method, by using the geometrical property(such as the location and orientation of stick-like features) interpreted in the obtained sketch map, a geometrical kernel function is adaptively constructed and combined with a ratio-based distance to form a patch-based metric that is valid for SAR image. Due to the usage of the kernel function implying the local geometrical property, the obtained metric becomes effective on the search of local homogeneous region. Moreover, by using the local stationary assumption that is widely used in image processing, classical region growing method is applied to determine the local maximal homogeneous region. Finally, maximum likelihood(ML) rule is adopted within the obtained region to estimate the true value of SAR image. Since the local homogeneous region was obtained by using the metric containing the sketch-based geometrical kernel, not only the speckle is strongly suppressed by the designed method, but also the stick-like features(e.g. edges and lines) are well preserved in the results.(3) Due to the ignorance of the structural information contained in SAR image, the same operation was usually used to estimate the true value of each pixel in classical filtering methods. However, owing to the existence of leading scatters in the resolution cell containing the stick-like features(e.g. edges and lines), the condition of full-developed speckle becomes invalid. This means, by using the operations designed for the regions containing non-structural information, the details contained in these regions would be smeared. Thus, given the capability of SAR sketch map on representing the stick-like features in SAR image, all pixels of a SAR image are classified into the directional pixel and the non-directional pixel in Chapter 4. For the directional pixel, due to the local significant directionality, a geometric-structural block(GB) is designed according to the local orientation and a GB based Non-local mean method is designed to estimate the true value. For the non-directional pixel, by designing a distribution based metric to measure the pixels’ similarity, an adaptive region(AN) based filter is adopted to calculate the underlying value. Since a block-wised method is used in GB based Non-local mean method, overlapping is inevitable and a distribution based weighting method is adopted to obtain the true value of the overlapped pixels. Due to the SAR sketch map based fusion of GB based NLM filter and AN based filter, not only the speckle is strongly suppressed in the result, but also some more details are well preserved, especially for some weak features and point-like target.(4) For classical adaptive region based SAR speckle reduction method, the adaptive region of each pixel is formed via the similarity between the concerned pixel and its neighbor pixels under a certain metric. However, for the pixels nearby the stick-like features(e.g. edges and lines), a stronger correlation exists along the local orientation than across the orientation; while correlation is decreasing with the distance from the center for point-like features, whose rate is inversely proportional to the size of the point-like feature. It means that local geometrical relationship among the pixels should be carefully considered in the metric. Inspired by the steerable kernel method designed for optical images, a SAR-steerable kernel method is proposed in Chapter 5 to interpret the local geometrical relationship and adopted for SAR speckle reduction. In this method, by analyzing the edge detectors valid for SAR image, a gradient operator is designed to calculate the gradient vector in SAR image. Then, with the designed gradient operator, a SAR-steerable kernel is constructed adaptively for each pixel and combined with a ratio-based metric to determine the local adaptive region of the pixel. Finally, with the obtained region, ML rule is adopted to estimate the true value of the pixel. As the SAR-steerable kernel is constructed with the parameters calculated from local gradient matrix, it is more effective to interpret the local geometrical property. The experimental results demonstrate that, due to the usage of SAR-steerable kernel in determining the local adaptive region, not only the speckle is largely reduced, but also the resolution of the details is well preserved in the results of our method. Moreover, the proposed method is of a better mean preservation.(5) Recently, non-local mean based filtering method has become one of the hot topics in image processing. However, in this kind of methods, the pixels’ similarity is always computed via their neighborhoods of fixed support. This is not good for the recovery of stick-like features(e.g. edges and lines) and point-like features which are always being smeared in the results. It is because that there is a stronger geometric correlation in the neighborhood containing the edge, lines and points. Thus, local geometric should be carefully explored and exploited to improve the robustness of the neighbor-based metric. In Chapter 6, considering the directional preference of the exponential weighting method used in the gradient method, a new weighting method that is of non-directional preference is proposed to calculate the local gradient value. Moreover, due to the diversity of scale information in SAR images, a multi-scale fusion based SAR-steerable kernel method is proposed and applied in SAR non-local mean method for SAR speckle reduction. In this method, by combining the obtained SAR-steerable kernel function with the speckle’s ratio based metric, the similarity between the center pixel and the pixels contained in the search window is computed, while weighted maximum likelihood method is adopted to calculate the center pixel’s value. The experimental results demonstrate that, due to the usage of SAR-steerable kernel based metric to compute the weight, the details of SAR image is well preserved while the speckle was reduced largely, especially for smaller targets.