Micro-Motion Classification for Radar Targets Based on Time-Frequency Semantic Structure

Spatial targets exist various micro-motions in flight. Micro-motion classification is important for distinguishing ballistic target types and parameter estimation. Most deep learning methods directly use unprocessed micro-Doppler spectrogram, which includes extensive irrelevant regions and is easily influenced by noise. To solve these questions, we introduce semantic structure for micro-motion classification. The semantic structure of time-frequency spectrogram is first extracted through multi-resolution saliency filtering and histogram-based filtering, which effectively eliminates noise while preserving critical motion features. With these informative motion features, high performance micro-motion classification is achieved by using deep network. Experiments show that this method improves recognition accuracy.