Abstract: As the electromagnetic spectrum environment becomes increasingly complex， enhancing the recognition capability of Interrupted Sampling Repeater Jamming （ISRJ）—a representative type of jamming characterized by strong concealment and dynamic behavior—has become crucial for ensuring radar detection performance. ISRJ is an intra-pulse coherent jamming method realized via partial sampling， storage， and coherent retransmission of radar signals. Existing recognition methods predominantly treat ISRJ as a single category， with limited modeling capacity and insufficient parameter generalization， thereby restricting their effectiveness in complex electromagnetic environments. To address this limitation， this study proposes an integrated recognition framework that synergistically combines time-frequency feature dimensionality reduction with a lightweight Vision Transformer architecture. The framework first employs a Kronecker separable dictionary learning algorithm based on closed-form Alternating Least Squares （ALS） for structure-preserving reduction of time-frequency representations. This approach effectively retains key discriminative features of ISRJ， including “intermittency” and “striped structures”， while substantially reducing computational cost and memory overhead. Subsequently， a lightweight Vision Transformer that integrates local convolutional priors and global self-attention mechanisms is developed to balance fine-grained feature representation with long-range dependency modeling， thereby improving robustness to parameter variations. The inference speed surpasses that of both standard Transformers and typical lightweight CNNs， with particularly pronounced advantages in multi-class ISRJ recognition under low Jamming-to-Noise Ratio （JNR） conditions， demonstrating improved efficiency and practical deployment potential.