针对传统基于统计理论的海面小目标检测方法在复杂海面环境中性能不高的问题,该文提出了一种改进的检测方法.首先通过分析海杂波和目标回波的特征,将检测问题转化为特征空间的分类任务.鉴于海面小目标样本数量有限,存在样本不平衡的问题,该文引入了一种基于梯度惩罚的沃瑟斯坦生成对抗网络(Wasserstein Generative Adversarial Network with Gradient Penalty,WGAN-GP)来增强目标数据,从而在数量上平衡目标样本与海杂波样本.同时,对原始WGAN-GP网络的损失函数进行了改进,引入相位损失以确保生成数据能够反映真实数据的相位信息.基于这些数据,进一步提取了生成目标和海杂波的高维特征,并将其送入卷积神经网络(Convolutional Neural Network,CNN)进行训练.为了应对高维特征空间中虚警概率难以控制的问题,对CNN算法进行了改进,通过设置Softmax分类器的阈值,实现了虚警概率可控.最后,借助公开的IPIX雷达数据集进行实验验证,所提的WGAN-GP-CNN检测器在积累时间为1.024 s,虚警概率为0.001时,平均检测概率达到0.8683,具有良好的检测效果.
Small Float Target Detection in Sea Clutter Based on WGAN-GP-CNN
This paper presents an improved method for detecting small targets on the sea surface,aiming to overcome the performance limitations of statistical-based approaches in complex sea environments.The detection problem is trans-formed into a classification problem in a feature space by extracting features from both sea clutter and target echoes.Due to the limited number of target samples on the sea surface,causing sample imbalance,this paper proposes a target data augmentation model based on Wasserstein generative adversarial network with gradient penalty(WGAN-GP).This model balances the target and sea clutter samples.The phase loss is designed to improve the overall loss function of the WGAN-GP,encouraging the generated samples to accurately learn and reproduce the phase characteristics exhibited by real data.Then,the high-dimensional features of the target and sea clutter are extracted and trained using a convolutional neural network(CNN).Next,because it is difficult to control the false alarm rate in a high-dimensional feature space,the CNN is improved by setting the threshold of the Softmax classifier,and the false alarm probability is controlled.Fi-nally,experimental validation is performed with the assistance of the publicly available IPIX radar dataset.The WGAN-GP-CNN detector proposed in this paper achieves an average detection probability of 0.8683 at an accumulation time of 1.024 s and a false alarm probability of 0.001,which provides good detection results.
NETL
NSTL
万方数据
