基于统计和纹理特征的SAS图像SVM分割研究  

Study on SAS image segmentation using SVM based on statistical and texture features

作　　者：陈强[1] 田杰[1] 黄海宁[1] 张春华[1] 

机构地区：[1]中国科学院声学所,北京100190

出　　处：《仪器仪表学报》2013年第6期1413-1420,共8页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(11174313)资助项目

摘　　要：合成孔径声呐图像可以有效反映海底的地形、地貌和底质等情况,但是单幅SAS图像通常对应一片较大的区域,需要按照某种性质将不同性质的区域分割开来,以有利于下一步的图像分析以及目标检测和识别。研究发现,不同底质区域的SAS图像具有不同的统计和纹理特征,选取灰度直方图的均值、标准差、峰度等统计特性和灰度共生矩阵的能量、相关性、对比度、熵值等纹理特性用以描述SAS图像的不同区域。将选取的特征作为SVM的训练特征,进而得到SVM分类器,用于SAS图像分割。实验结果表明,SVM算法可以很好地对SAS图像进行区域分割。Synthetic aperture sonar （SAS） images can effectively describe the topography,geomorphology and substrate of seabed;however, one single SAS image usually corresponds to a larger area;so it is necessary to segment the SAS image in- to different regions according to certain property ,which benefits further analyzing the image ,and detecting and identifying the target. Study found that SAS images of different substrates have different statistical and texture features;in this paper, the statistical properties,such as the mean,standard deviation and kurtosis of the grey level histogram, as well as the tex- ture features,such as the energy,correlation,contrast and entropy of the grey level co-occurrence matrix are selected and used to describe different regions of the SAS image. These selected features are used as the support vector machine （SVM） training characteristics and the classifier is obtained for the SAS image segmentation. The experiment results show that the proposed SVM algorithm is a good segmentation method for the region segmentation of SAS image.

关 键 词：合成孔径声呐 统计特征 纹理特征 图像分割 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]