摘要
以机器学习为主的雷达目标识别模型性能由模型与数据共同决定。当前雷达目标识别评估依赖于准确性评估指标,缺乏数据质量对识别性能影响的评估指标。数据可分性描述了属于不同类别样本的混合程度。数据可分性指标独立于模型识别过程,将其引入识别评估过程,可以量化数据识别难度,预先为识别结果提供评判基准。因此该文基于率失真理论提出一种数据可分性度量,通过仿真数据验证所提度量能够衡量多维高斯分布数据的可分性优劣。进一步结合高斯混合模型,设计的度量方法能够突破率失真函数的局限性,捕捉数据局部特性,提高对数据整体可分性的评估精度。接着将所提度量应用于实测数据识别难度评估中,验证了其与平均识别率的强相关性。而在卷积神经网络模块效能评估实验中,首先在测试阶段量化分析了各卷积模块提取特征的可分性变化趋势,进一步在训练阶段将所提度量作为特征可分性损失参与网络优化过程,引导网络提取更可分的特征,该文从特征可分性角度为神经网络识别性能的评估与提升提供新思路。
The performance of machine learning-based radar target recognition models is determined by the respective model and data to be analyzed.Currently,radar target recognition performance evaluation is based on accuracy metrics,but this method does not include the evaluation metrics regarding the impact of data quality on recognition performance.Data separability describes the degree of mixture of samples from different categories.Furthermore,the data separability metric is independent of the model recognition process.By incorporating it into the recognition evaluation process,recognition difficulty can be quantified,and a benchmark for recognition results can be provided in advance.Therefore,in this paper,we propose a data separability metric based on the rate-distortion theory.Extensive experiments on multiple simulated datasets demonstrated that the proposed metric can compare the separability of multivariate Gaussian datasets.Furthermore,by combining it with the Gaussian mixture model,the designed metric method could overcome the limitation of the rate-distortion function,capture the data’s local separable characteristics,and improve the evaluation accuracy of the overall data separability.Subsequently,we applied the proposed metric to evaluate the recognition difficulty in real datasets,the results of which validated its strong correlation with average recognition accuracy.In the experiments on evaluating the effectiveness of convolutional neural network modules,we first quantified and analyzed the separability trend of the feature extracted by each module during the testing phase.Further,we incorporated the proposed metric as a feature separability loss during the training phase to participate in the network optimization process,guiding the network to extract a more separable feature.This paper provides a new perspective for evaluating and improving the neural network recognition performance in terms of feature separability.
作者
姜卫东
薛玲艳
张新禹
JIANG Weidong;XUE Lingyan;ZHANG Xinyu(College of Electronic Science and Technology,National University of Defense Technology,Changsha 410073,China)
出处
《雷达学报（中英文）》
EI
CSCD
北大核心
2023年第4期860-881,共22页
Journal of Radars
基金
国家自然科学基金(61921001)。
关键词
机器学习
雷达目标识别评估
率失真理论
识别率
数据可分性度量
Machine learning
Radar target recognition evaluation
Rate distortion theory
Recognition accuracy
Data separability metric
