针对深度学习模型易出现灾难性遗忘的关键难点,提出了一种基于原型对比的合成孔径雷达(synthetic aperture radar,SAR)图像增量小样本目标检测算法-InFSAR(prototype contrast based incremental few-shot SAR object detection)。首先...针对深度学习模型易出现灾难性遗忘的关键难点,提出了一种基于原型对比的合成孔径雷达(synthetic aperture radar,SAR)图像增量小样本目标检测算法-InFSAR(prototype contrast based incremental few-shot SAR object detection)。首先,采用基础数据集对检测器进行预训练,以构建初步的特征提取能力;其次,设计一种类原型表征生成模块,以构建一组能够代表数据内在特征的类原型。在增量学习阶段,设计一种混合类原型对比编码模块,以有效学习新类别与基础类别之间的区分性特征。此外,为缓解灾难性遗忘问题,引入类原型校准策略,使模型在类原型上的预测分布逐步逼近真实分布,从而保持对基础类别识别的稳定性。在小样本目标检测数据集SRSDD-v1.0上的实验结果表明,在5-shot设置下,InFSAR对船舶细粒度目标的检测精度达到46.5%。同时,该方法能够在无需访问基础类训练数据的情况下,实现对少量标注新类别的增量检测与识别。展开更多
Affected by the insufficient information of single baseline observation data,the three-stage method assumes the Ground-to-Volume Ratio(GVR)to be zero so as to invert the vegetation height.However,this assumption intro...Affected by the insufficient information of single baseline observation data,the three-stage method assumes the Ground-to-Volume Ratio(GVR)to be zero so as to invert the vegetation height.However,this assumption introduces much biases into the parameter estimates which greatly limits the accuracy of the vegetation height inversion.Multi-baseline observation can provide redundant information and is helpful for the inversion of GVR.Nevertheless,the similar model parameter values in a multi-baseline model often lead to ill-posed problems and reduce the inversion accuracy of conventional algorithm.To this end,we propose a new step-by-step inversion method applied to the multi-baseline observations.Firstly,an adjustment inversion model is constructed by using multi-baseline volume scattering dominant polarization data,and the regularized estimates of model parameters are obtained by regularization method.Then,the reliable estimates of GVR are determined by the MSE(mean square error)analysis of each regularized parameter estimation.Secondly,the estimated GVR is used to extracts the pure volume coherence,and then the vegetation height parameter is inverted from the pure volume coherence by least squares estimation.The experimental results show that the new method can improve the vegetation height inversion result effectively.The inversion accuracy is improved by 26%with respect to the three-stage method and the conventional solution of multi-baseline.All of these have demonstrated the feasibility and effectiveness of the new method.展开更多
文摘针对深度学习模型易出现灾难性遗忘的关键难点,提出了一种基于原型对比的合成孔径雷达(synthetic aperture radar,SAR)图像增量小样本目标检测算法-InFSAR(prototype contrast based incremental few-shot SAR object detection)。首先,采用基础数据集对检测器进行预训练,以构建初步的特征提取能力;其次,设计一种类原型表征生成模块,以构建一组能够代表数据内在特征的类原型。在增量学习阶段,设计一种混合类原型对比编码模块,以有效学习新类别与基础类别之间的区分性特征。此外,为缓解灾难性遗忘问题,引入类原型校准策略,使模型在类原型上的预测分布逐步逼近真实分布,从而保持对基础类别识别的稳定性。在小样本目标检测数据集SRSDD-v1.0上的实验结果表明,在5-shot设置下,InFSAR对船舶细粒度目标的检测精度达到46.5%。同时,该方法能够在无需访问基础类训练数据的情况下,实现对少量标注新类别的增量检测与识别。
基金National Natural Science Foundation of China(No.42104025)China Postdoctoral Science Foundation(No.2021M702509)+3 种基金Natural Resources Sciences and Technology Project of Hunan Province(No.2022-07)Surveying and Mapping Basic Research Foundation of Key Laboratory of Geospace Environment and Geodesy,Ministry of Education(No.20-01-04)Natural Science Foundation of Hunan Province(No.2024JJ5144)Open Fund of Hunan International Scientific and Technological Innovation Cooperation Base of Advanced Construction and Maintenance Technology of Highway(Changsha University of Science&Technology,No.kfj190805).
文摘Affected by the insufficient information of single baseline observation data,the three-stage method assumes the Ground-to-Volume Ratio(GVR)to be zero so as to invert the vegetation height.However,this assumption introduces much biases into the parameter estimates which greatly limits the accuracy of the vegetation height inversion.Multi-baseline observation can provide redundant information and is helpful for the inversion of GVR.Nevertheless,the similar model parameter values in a multi-baseline model often lead to ill-posed problems and reduce the inversion accuracy of conventional algorithm.To this end,we propose a new step-by-step inversion method applied to the multi-baseline observations.Firstly,an adjustment inversion model is constructed by using multi-baseline volume scattering dominant polarization data,and the regularized estimates of model parameters are obtained by regularization method.Then,the reliable estimates of GVR are determined by the MSE(mean square error)analysis of each regularized parameter estimation.Secondly,the estimated GVR is used to extracts the pure volume coherence,and then the vegetation height parameter is inverted from the pure volume coherence by least squares estimation.The experimental results show that the new method can improve the vegetation height inversion result effectively.The inversion accuracy is improved by 26%with respect to the three-stage method and the conventional solution of multi-baseline.All of these have demonstrated the feasibility and effectiveness of the new method.
