This paper presents a robust radar work mode recognition framework combining multi-scale time-frequency analysis with vision transformer architecture. Our approach introduces novel multi-scale time-frequency representations that encode pulse-level, intra-group, and inter-group signal characteristics through RGB channel fusion, effectively capturing complementary waveform features. The modified Vision Transformer architecture incorporates dual-level routing attention (DRA) to dynamically focus computational resources on critical time-frequency regions. Comprehensive evaluations demonstrate significant performance improvements, with the proposed method attaining 85.7% recognition accuracy at –6 dB SNR and over 99% accuracy beyond 6 dB SNR. These advancements provide a viable solution for modern radar threat identification under challenging electromagnetic conditions.