Making full and effective use of target polarization information from High Resolution Range Profile (HRRP) is a hot issue for improving the recognition performance of maritime surveillance radar. A HRRP database with seven maritime targets classes from various aspect angles is established, on which thirty-nine features from four categories are defined. A novel feature selection method based on the Normalized Mutual Information (NMI) and Simulated Annealing (SA) algorithm is presented, named as NMI-SA. The effectiveness of the NMI-SA is proved by comparison with three other methods using HRRP dataset and eight from UCI machine learning repository. Finally, the NMI-SA is applied to the HRRP dataset to find twenty-five high discriminant and low redundancy features.%充分、有效地利用目标全极化HRRP的特征信息是提高对海雷达目标识别率的研究热点之一。该文利用CST软件仿真建立了7类海上目标在不同方位角下的全极化HRRP数据库。在此基础上,提取了4类共39个特征。提出一种基于归一化互信息(NMI)并利用模拟退火(SA)算法进行优化的全局最优特征选择算法,并命名为NMI-SA。基于HRRP数据集以及9个UCI数据集,利用k-近邻分类器将该算法与另外3种常用的特征选择算法进行对比,结果表明新算法选择的特征具有良好的可分性和较低的冗余度,最终用于分类时的正确率总体优于其余3种算法。最后,用该算法对全极化HRRP的39个特征进行重点分析,选择出25个辨别力强、冗余度低的特征。
展开▼