基于Radon变换的SAR图像目标统计不变特征提取研究

The Research of the Statistically Invariant Features Extraction for SAR Target Image Based on Radon Transform

SAR目标图像敏感于目标方位的特性为SAR解译带来了困难。通过对目标散射机制的分析,认为一定方位角变化范围内的目标图像只是在图像域发生了旋转和平移,其灰度统计分布是不变的,利用X波段的实测车辆目标数据,通过Radon变换提取图像64个投影方向上序列的各阶矩以及序列的极值、宽度等灰度统计值构造特征矩阵,然后对特征矩阵进行旋转和平移校正得到统计不变特征。结果表明:该特征在5°方位角范围内具有70%以上的稳定性,因此可以用来合并目标不同方位角的信号。

It is very hard for SAR target recognition because of the characteristic that the SAR target image is strongly dependent upon the target's azimuth angle. Theoretically, SAR image is dominated by reflection from corner reflector, and the back scattering of the corner reflector can generate persistent scattering over definite azimuth angle range, so the target SAR images can be invariant for this range. In this paper's view, SAR images in different azimuth angles can be regarded as the rotation and shift transform in image domain, so the statistical distribution of pixels' gray-scale should be invariant. By the operations performed on the Radon transform of target chips, the 64 projections of every target chip are taken. The mean, width and central moments of 2~6 orders for each projection are calculated to construct the feature matrix. The feature matrix is statistically invariant to changes in rotation and shift. Utilizing MSTAR measured vehicle target data, this paper extracts the feature of target image based on the Radon transform and analyzes the stability of the feature by computing the correlation matching metric between SAR images in different azimuth angles. The result support our conclusion of invariance of the feature, because it proves that the feature has over 70% stability in 5° azimuth angle range. So this algorithm can be used to incorporate target signatures at as many orientations as possible.