Hierarchical cross attention model for multi-modal HRRP recognition

The ease of processing, small data size, and rich target structural information associated with High-Resolution Range Profile (HRRP) have consistently positioned it as a research focal point in the Radar Automatic Target Recognition (RATR) field. However, most existing methods primarily focus on a single time domain (amplitude) feature, neglecting the time -dependent and multi-domain features of HRRP. In this regard, this paper proposes a multimodal hierarchical cross-attention model that combines recurrent and co-attention neural networks. Initially, the time and frequency domain representations are processed using a set of bidirectional recurrent neural network layers with self-attention, which can transform the dual-domain information of HRRP into fixed-dimensional embeddings. Additionally, the incorporation of co-attention layers integrates contextual information from both modalities, with the layer attempting to weigh the utterance level embeddings relevant to the task of HRRP target recognition. Finally, the neural network parameters in the time domain layer, frequency domain layer, and multimodal co-attention layer are hierarchically trained for the HRRP target classification task. The experimental results indicate that the algorithm can effectively learn the multi-domain temporal correlations of HRRP, thereby enhancing target recognition performance.