Synthetic aperture radar (SAR) automatic target recognition (ATR) technology has seen significant advancements. Despite these advancements, the majority of research still operates under the closed-set assumption, wherein all test samples belong to classes seen during the training phase. In real-world applications, however, it is common to encounter targets not previously seen during training, posing a significant challenge to the existing methods. Ideally, an ATR system should not only accurately identify known target classes but also effectively reject those belonging to unknown classes, giving rise to the concept of open set recognition (OSR). To address this challenge, we propose a novel approach that leverages the unique capabilities of the Capsule Network and the Kullback-Leibler divergence (KLD) to distinguish unknown classes. This method begins by deeply mining the features of SAR targets using the Capsule Network and enhancing the separability between different features through a specially designed loss function. Subsequently, the KLD of features between a testing sample and the center of each known class is calculated. If the testing sample exhibits a significantly larger KLD compared to all known classes, it is classified as an unknown target. The experimental results of the SAR-ACD dataset demonstrate that our method can maintain a correct identification rate of over 95% for known classes while effectively recognizing unknown classes. Compared to existing techniques, our method exhibits significant improvements.

Existing Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) systems, when confronted with novel, unknown target classes in real-world scenarios, tend to forcibly categorize them into one of the known classes used during training. This approach can lead to immeasurable losses in practical applications. Consequently, achieving open-set recognition (OSR) for SAR targets is of paramount importance, as it requires both effective discrimination among known classes and the ability to identify unknown classes. Building upon previous research, this study proposes a SAR OSR method based on the generation of feature instances for unknown classes. The proposed method employs the Capsule Network as the feature extraction network, incorporating a self-attention mechanism to enhance the effectiveness and robustness of the extracted features. By randomly combining features, the method simulates and generates feature instances of unknown class targets using the features of known class targets. Both known class features and unknown class feature instances are then input into the network for training, resulting in a K+1 closed-set classifier that effectively addresses the OSR problem. Comparative experiments with other methods demonstrate that our approach exhibits superior performance.

Open-set synthetic aperture radar (SAR) automatic target recognition must not only classify familiar targets accurately, but also reject target categories that are absent from training. This task is particularly challenging because aspect-dependent scattering produces substantial and often multi-modal intra-class variation, while the unknown categories available for validation cannot fully represent the open space encountered at deployment. To address these coupled difficulties, we propose the Dynamic Adaptive Multi-scale ASC Network (DAMS-ASCNet), a unified representation–distribution–decision framework for open-set SAR recognition. First, a dynamic adaptive multi-scale angle-sensitive convolution module learns sample-dependent directional and scale responses, enabling the feature extractor to preserve informative scattering variations while suppressing irrelevant disturbances. Second, multiple prototypes are introduced to model the aspect-induced submodes of each known class, providing a more faithful class representation than a single-center description. Third, open-set decisions are performed using class-conditional prototype similarities and class-conditional calibration, with rejection thresholds estimated exclusively from validation data. The revised evaluation separates the effects of representation, training objective, prototype modeling, and calibration, and reports threshold-dependent operating-point metrics separately from threshold-independent score-discrimination metrics. Experiments on the MSTAR open-set benchmark show that, on the original open-set split, the proposed framework with validation-only class-conditional calibration achieves $80.33\pm 2.96\%$ open-set accuracy and $81.27\pm 5.48\%$ unknown recall over three seeds; additional held-out class compositions reveal substantial sensitivity to the choice of unknown classes. Controlled ablations further show that the matched training objective materially affects the calibrated operating point, while the architecture-level advantage over an ordinary-convolution counterpart is modest; excessive feature compactness can also be detrimental to open-set separability.