针对合成孔径雷达(SAR)图像中普遍存在具有相干斑噪声、流冰区域浮冰接触紧密及小碎冰较多等背景复杂问题,提出了一种基于背景抑制显著性检测的SAR流冰分离算法。首先基于图像的显著性检测,通过学习随机森林回归量得到初步显著性图;然后通过超像素构建图像区域并进行离散傅里叶变换,提取区域频域特征并计算卡方距离;之后对边界背景进行抑制,生成背景抑制模块图;最后将两阶段的图融合得到增强显著性图。在SAR海冰数据集上对所提算法、7种显著性算法及3种海冰分割方法进行对比实验。结果表明,所提算法可以有效检测出孤立浮冰,抑制背景区域。

In the synthetic aperture radar (SAR) images, there are many complex background problems, such as speckle noise, close contact of floating ice in drift-ice area, and more small pieces of ice. Here, we describe a SAR flow ice separation algorithm based on background-suppression saliency detection. This algorithm obtains a preliminary saliency map by learning the random forest regression based on the saliency detection of images. Then the image region is constructed by using super pixels. Additionally, discrete Fourier transformation is performed, after which the frequency domain features of the region are extracted, and the chi-squared distance is calculated. The proposed algorithm then suppresses the boundary background to generate a background-suppression module diagram, followed by fusion of the two-stage graph to obtain the enhanced saliency graph. We evaluate the proposed algorithm and compare its performance against seven saliency algorithms and three sea-ice-segmentation methods using a SAR sea-ice dataset. The results show that the proposed algorithm can effectively detect isolated ice floes and suppress background areas.