多特征的光学遥感图像多目标识别算法  

Research on multi-feature based multi-target recognition algorithm for optical remote sensing image

作　　者：姬晓飞[1] 秦宁丽 刘洋[1] JI Xiaofei QIN Ningli LIU Yang(School of Automation, Shenyang Aerospace University, Shenyang 110136, China Beijing GuoDianTong Network Technology Co Ltd, Beijing 100070, China)

机构地区：[1]沈阳航空航天大学自动化学院,辽宁沈阳110136 [2]北京国电通网络技术有限公司,北京100070

出　　处：《智能系统学报》2016年第5期655-662,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61103123);辽宁省高等学校优秀人才支持计划项目(LJQ214018);辽宁省自然科学基金项目(2015020101)

摘　　要：基于单一特征的光学遥感图像多目标分类识别存在准确性较差的问题,提出一种新的基于多特征决策级融合的多目标分类识别算法。首先对光学遥感图像目标提取3种能够同时满足平移、旋转和尺度不变性的特征:可以描述局部和全局分布特性的分层Bo F-SIFT特征,描述目标边缘轮廓点信息的改进后的SC形状特征,对图像中较大目标识别较好的Hu不变矩特征;其次采用基于径向基核函数的一对一支持向量机算法分别获得3种特征的目标识别概率,并设计了一种多特征决策级加权融合的策略实现对多目标的分类。经多次实验验证该算法对光学遥感图像多目标具有较好的分类识别性能,达到了93.52%的正确识别率。A novel multi-feature decision level fusion recognition algorithm is proposed to solve the problem of poor levels of accuracy with single feature based multi-target classification of optical remote sensing images. Firstly, three kinds of features which can not only meet translation, rotation, and scale invariance are extracted. One is the hier- archical BoF-SIFT feature which can simultaneously describe local and global distributions. Another is the improved shape context feature which is used to describe the target edge contour point information. The other one is Hu mo- ment invariants which gives better levels of recognition performance for large targets. Secondly, the recognition probabilities of these features are obtained using a one versus one support vector machine based on a radial basis function. Thirdly a strategy for multi-feature decision level fusion is designed. A large number of experiments show that the algorithm for multi-target classification of optical remote sensing images performs better with the recognition rate of targets reaching 93.52%.

关 键 词：光学遥感图像 多特征的决策级融合 分层的BoF-SIFT特征 SC形状特征 Hu不变矩特征 支持向量机 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]