Radar Automatic Target Open Set Recognition Based on Cluster Boundary Detection

Radar automatic target recognition has developed significantly in recent years. Previous works mainly aim at addressing the closed set recognition problem, whose test set and training set have the same categories of the targets. However, radar easily encounters the target categories outside the training set in the practical application, and closed set recognition could cause incorrect classification problem under the circumstance. Therefore, open set recognition (OSR) is proposed to solve this problem, which can classify known targets and identify unknown targets simultaneously. Currently, there are lots of works focusing on optimizing feature representation learning via complex deep models, while existing decision boundary modeling methods—though increasingly explored—mostly rely on Euclidean distance metrics or simple thresholding, struggling to adapt to the complex distribution of high-dimensional features. In fact, constructing precise closed decision boundaries in the feature space is equally critical for improving OSR performance. Guided by this intention, this paper proposes a novel cluster boundary detection algorithm based on vector angle distribution, which can accurately detect specific boundary points of clusters. Once the closed boundary is determined, distinguishing between known and unknown targets becomes straightforward. Finally, extensive experiments on measured radar high-resolution range profile (HRRP) data verify the validity and effectiveness of the proposed method.