高分辨距离像(HRRP)分类是对雷达复杂目标分类的一种重要方法。标准的一对一超球面SVM多值分类方法需要训练k(k-1)个子分类器,计算量大、训练时间长,并且存在决策盲区,不适宜用来进行HRRP目标识别。为了减少分类器数量,提高训练速度,文中根据超球面的几何特征引入了一种“倒数对称”的一维隶属度,构造了模糊超球面SVM分类器,该方法仅需训练k(k-1)/2个子分类器,既提高了训练速度又解决了决策盲区,HRRP实测数据识别实验表明了该方法的有效性。

High resolution range profile (HRRP) classification is an important method for radar complex target classification. Since standard one-against-one hypersphere support vector machine (SVM) has the defects of large computation, long training time for its k(k-1) sub-classifiers, and, decision bland area, it is not fit for HRRP target recognition. In order to reduce the number of classifiers in the one-against-one multi-class, a new one-dimensional membership function based on geometry feature named "reciprocal symmetry" has been defined, and the corresponding fuzzy hypersphere SVM has been given. This new method only needs k(k-1) /2 sub-classifiers, it not only improves the training speed, but also clears away the decision bland area. The HRRP real data experimental results show that this algorithm has better HRRP classification performance.