基于虚警可控梯度提升树的海面小目标检测

Sea-surface small target detection based on false-alarm-controllable gradient boosting decision tree

高维特征检测是提升海面小目标探测性能的一种有效途径,其主要难点在于高维空间分类器设计.本文提出一种基于虚警可控梯度提升树(Gradient Boosting Decision Tree,GBDT)的特征检测方法.首先,从一维长时观测向量中,提取时域、频域、时频域等多个特征,构建高维特征向量,从而将检测问题转换为二分类问题.其次,通过仿真含目标回波,解决两类训练样本非均衡的问题.然后,引入GBDT算法,将高维特征向量凝聚为一维概率预测值,并以预测值作为检测统计量,解决二分类器难以控制虚警的问题.最后,采用IPIX实测数据验证,结果表明:本文所提的检测器充分利用了高维特征的全部信息,性能平均提升13%以上.

At present,high-dimensional(HD)feature detection is an effective approach to improve the detection performance of sea-surface small targets.The main difficulty lies in the design of classifier in high-dimensional space.Therefore,a feature detection approach based on false-alarm-controllable gradient boosting decision tree(GBDT)is proposed in this paper.First,multiple features are extracted from the 1D long-term observation vector in time,frequency,time-frequency domains to construct an HD feature vector.In this way,the detection problem is converted into a binary classification problem.Second,two types of balanced training samples are solved by simulating returns with target.Third,GBDT algorithm is introduced to condense the HD feature vector into 1D predicted value in probability.The predicted value is used as detection statistics to solve the problem of uncontrollable false alarm rate perplexed the binary classifier.Finally,experimental results are verified by IPIX measured data,which show that the proposed detector can make full use of all the information from the HD characteristics,and the performance is improved by over 13%.