Hierarchical State Space Representation Learning for Vehicle Classification in ISAR Imagery

Due to the advantages of all-weather, all-day and strong penetration, and so forth, millimeter wave radar has been widely used in intelligent transportation systems. Millimeter wave based ISAR image can provide rich information for environment sensing in intelligent transportation systems. Target Classification is one of the important means of ISAR image interpretation. Therefore, this paper proposes a novel method called hierarchical state space representation learning (HSSRL) to achieve robust vehicle target classification in ISAR imagery. Specifically, a local feature extraction module based on variable kernel convolution is first established to realize feature dimension reduction and local discriminative feature extraction from ISAR target image. Secondly, a global feature extraction module integrating bidirectional selective state space model and attention mechanism is developed to achieve long-distance global dependency modeling of ISAR images. In order to effectively integrate useful information, the proposed method establishes a global-local interactive learning mechanism, which can further improve feature discriminability with the help of skip connection and attention mechanism. And then, two fully connected layers are introduced to achieve feature aggregation and target identity reasoning. Finally, we introduce an improved cross-entropy loss to optimize the model, attempting to construct a feature space with intra-class compactness and inter-class dispersion, thereby achieving efficient vehicle target classification. A series of evaluation experiments on the ISAR vehicle target dataset demonstrate the effectiveness and superiority of the proposed method.