SAR Image Segmentation Techniques Based On Statistical Model

Synthetic Aperture Radar (SAR) image segmentation is one of the key techniques for SAR image automatic interpretation. Due to the multiplicative speckle noise included in SAR images, most traditional segmentation methods, which are designed for additive noise, are not applicable to SAR images. Since the statistical characteristics of SAR images are quite different from that of common optical images, some statistical information based image segmentation techniques, which are more accordant to the nature of SAR images, have received increasing interests. This kind of techniques establishes certain statistical model, constructs the criterion with probability principles and segments images by optimizing such criterion. More importantly, these methods need not the process of despeckling and can even perform speckle filtering while segmenting images.This dissertation mainly investigates several SAR image segmentation methods based on statistical model. The main research works are listed as follows:(1) The despeckle filters and edge detectors are studied respectively. First, traditional square local windows are replaced by non-square ones, which can preserve image edges and details in despeckling. Experiments on real SAR images filtering demonstrate the importance of the non-square local windows, as well as the effects of different choices of local windows on despeckling. Further, the ratio-of-averages (ROA) detector, being with the characteristic of constant false alarm rate, is introduced. Since it can perform edge detection in SAR images without despeckling, such scheme shows great advantage over gradient-based edge detectors. Besides, the effects of the false alarm rate and the size of local windows on edge detection are also discussed.(2) In order to reduce the number of free parameters in segmentation, a SAR image segmentation technique based on minimum description length (MDL) principle is proposed. First, it establishes statistical models for logarithm version of SAR images. Then a criterion is constructed based on the MDL principle, with the supposition of the underlying piecewise constant images, which leads to the reconstructed images by optimizing such criterion. Finally, the histogram threshold is required as a postprocessing step to get perfect segmentation. The whole approach can automatically be carried out without parameters tuning, and is potential for practical applications.(3) The rest of paper focuses on SAR image segmentation based on level set, which has shown various merits on image segmentation. Taking into account the limitation that the geodesic active contour (GAC) model only uses edge information while the geodesic active region (GAR) model only uses region uniformity, a level set SAR image segmentation technique based on the ROA-G model is proposed. It combines the ROA edge information with region statistical characteristic of gamma probability density function (PDF), which has shown great superiority over both GAC model and GAR model. This claim is verified by segmentation experiments on both simulated and real SAR images.(4) Along with the increasing resolution of SAR images, the statistical model of high-resolution SAR images may no longer obey a single gamma PDF. For this reason, this paper presents a novel level set based high-resolution SAR image segmentation technique, which treats target region and background region with different statistical PDFs. Both theoretical analysis and simulation experiments indicate that for SAR images with different resolution or of different scenes, the“heavy-tailed”Fisher PDF is more suitable for describing the target region with strong reflectors, while the background region without strong reflectors could be still modeled by the gamma PDF. Indeed, the Fisher PDF does not belong to the exponential family of distribution, which makes the ML estimator be not applicable to its parameter estimation. Thus, three parameter estimation methods are also investigated to deal with the Fisher PDF.(5) Generally speaking, the parameter estimation for the Fisher PDF is time consuming. To alleviate such problem, a supervised SAR image segmentation approach based on level set is proposed, which just performs parameter estimation of properly chosen training data in original images before segmentation. It is unnecessary to re-estimate the PDF parameters during each step of curve iteration, so only once of parameter estimation is required in entire segmentation process, which consequently speeds up the final segmentation, especially for the Fisher PDF. Experimental results verify that this supervised segmentation approach achieves almost the same partition as the unsupervised one, while decreasing the running time effectively, if only the training data in the target region and the background region of images are properly chosen. It should be also noted that the saved time efficiency may be relevant to the complexity of parameter estimation of corresponding PDF, as well as the number of iterations of curve evolution.