Study On Target Detection In Synthetic Aperture Radar Images

Target detection in SAR images is an important step of the SAR imageinterpretation. This thesis focuses on several key problems of target detection in SARimages, including statistical characterization of the SAR image data, speckle reduction,and design of target detectors. The main work and innovations are summerized asfollows.1) An unsupervised estimation method of the equivalent number of looks (ENL) isproposed. Accurate description of the clutter distribution is the key for contant falsealarm rate (CFAR) detection and the ENL is an important parameter that characterizesthe statisitcal behavior of the SAR image data. The proposed method calculates the ENLby estimating the variance of the additive noise in the logorithmic SAR image alongwith a pre-established lookup table between the variance and the ENL. Compared withtraditional methods, the new method does not rely on the assumption that the imagecontains homogeneous areas; and it is easy to implement thus sutible for real-timeapplications.2) A generalized speckle reduction method for SAR images is proposed. Specklenoise is a major factor that affects the interpretation of SAR images and it causes falsealarms in the process of target detection. Starting from the polarimetric whitening filter(PWF), its equivalent formulation is derived, according to which a generalized specklereduction model is thus developed. In particular, an algorithm for reducing the specklein single-look SAR images is proposed based on the new model. In addition, forpreserving the power information of the output of the PWF, an optimal method isproposed. Experiments shows that the best speckle filtering effect is achieved by thePWF with the optimal power preservation.3) A new non-parametric target detector is proposed. The design of CFAR targetdetectors is one of the central focues in target detection. By combining two commonlyused non-parametric methods, the kernel density estimation (KDE) and the mean squareerror (MSE) distance, a new non-parametric test statistic is derived and used for targetdetection. In particular,CFAR detection is made possible in SAR images withmultiplicative noise. The main advantage of the new method is that it not only has a comparable performance to the optimal parametric detectors when the noise modelconforms to a given distribution, but also retains a strong robustness in stuationsotherwise.4) In order to solve the problem of inaccurate clutter estimation in the presense ofmutiple targets, an iterative censoring approach is proposed. By basing the outlierrejection in the next iteration on the detection result of the last iteration, a unifiedapproach of simultaneous detection of targets and rejection of outliers is realized. Thismethod can be successfully applied to both single-channel and multi-channel SARimages and is easy to implement.