图像分割是合成孔径雷达(SAR)图像自动目标识别应用中的基础性问题。定义并分析了运动、静止目标获取与识别(MSTAR)图像的矩特征,进而构造了多阈值处理策略,用于MSTAR图像的分割。首先研究了目标、阴影,以及背景区域统计特性,并确定了相应的数学模型描述,在此基础上给出了矩特征的定义,并分析了其基本特性。利用图像空间到矩特征空间的转换,显著增强了目标区域与阴影、背景区域的差异性,进而通过构造不同的阈值化规则,实现了MSTAR图像中目标、阴影和背景区域的分割。对MSTAR图像的处理结果表明,与恒虚警率(CFAR)、最大类间方差(OTSU)、模糊C均值(FCM)和马尔可夫随机场(MRF)等典型分割算法相比,本文算法不需进行噪声抑制,但在分割效果和鲁棒性等方面性能更好,同时,对多尺度、多目标MSTAR图像的分割也显示出良好的适应性。

Image segmentation is a fundamental step for SAR image based automatic target recognition (ATR). A method based on moment feature and multiple thresholds is proposed for moving and stationary target a cquisition and recognition (MSTAR) image segmentation. First, according to a comprehensive research of the statistics of MSTAR images, the mathematical description models for target regions, shadows, and background regions are constructed respectivedly. Then, the moment feature is defined, followed by the analysis of its basic properties. By transforming the image into moment feature space, the difference between target region and the other two types of regions is significantly enhanced. Finally, a strategy with multiple thresholds is constructed for the segmentation. The experiment results with MSTAR dataset indicate that the algorithm presented here has advantages not only on the noise robustness, but also on the segmentation effect, as well as the processing efficiency over the common-used methods, such as OTSU, fuzzy C-means(FCM), Markov random field(MRF), and constant false alarm rate (CFAR). Furthermore, this new method also performs well in the segmentation of MSTAR images with various scales and multiple targets.