基于多球体空间拓扑约束的雷达目标HRRP少样本开集识别方法OA

Radar target high-resolution range profile(HRRP)has significant application value in radar target recogni-tion due to its ability to effectively characterize the geometric structure and electromagnetic scattering properties of targets.In recent years,deep learning methods have been widely applied to HRRP-based target recognition tasks owing to their powerful feature representation capability.However,existing deep learning-based radar target recognition methods using HRRP suffer from performance degradation in few-shot open-set recognition(FSOSR)scenarios due to limited training samples and inherent inability to discriminate unknown classes.To address this issue,this paper proposes a radar target HRRP few-shot open-set recognition method based on multi-hypersphere spatial topological constraints.First,a meta-learn-ing framework is introduced in the training strategy.A large number of similar yet non-overlapping tasks are sampled from an auxiliary dataset with sufficient categories and abundant samples as training units to learn cross-task common knowl-edge,enabling the model to generalize rapidly on sample-scarce task datasets and alleviating few-shot overfitting.On this basis,a multi-hypersphere decision boundary modeling mechanism suitable for few-shot scenarios is designed.Each known class feature subspace is modeled by an individual hypersphere,forming a multi-hypersphere representation of the known-class feature space,which implicitly models the distribution of unknown classes.Meanwhile,adaptive margins are intro-duced between hyperspheres to maintain specific spatial topological relationships among known classes,thereby improving the refinement and robustness of the decision boundaries.Furthermore,a hard sample learning strategy is proposed.High-value samples are first selected through hard sample mining,after which a sample-wise weighting mechanism is employed to quantitatively characterize the hardness of samples.The model then performs targeted learning according to the assigned weights,enabling the extraction of finer discriminative features from informative hard samples and enhancing the model's capability to distinguish fine-grained unknown classes.Experimental results demonstrate that,in the 5-shot and 10-shot sce-narios,the proposed method achieves accuracy improvements of 6.17 percentage point and 2.94 percentage point,and AU-ROC improvements of 13.1 percentage point and 12.94 percentage point compared with the state-of-the-art methods,respec-tively,verifying the effectiveness and robustness of the proposed approach.In addition,the model is deployed on the Rock-chip RK3588 embedded artificial intelligence(AI)chip,achieving an inference latency of 2.197 ms and a power consump-tion of 2.25 W,which demonstrates its engineering feasibility.The proposed method is suitable for complex environments where unknown classes frequently appear,and can support high-resolution radar systems in the detection and recognition of non-cooperative targets.Moreover,due to its favorable engineering feasibility,the method is also applicable to scenarios with stringent real-time requirements and limited hardware resources,such as airborne platforms.