Recognizing the HRRP by Combining CNN and BiRNN With Attention Mechanism

Jinwei Wan, Bo Chen, Yingqi Liu, Yijun Yuan, Hongwei Liu, Lin Jin

Abstract

Radar Automatic Target Recognition (RATR) using High-Resolution Range Profiles (HRRP) is often hindered by sensitivities to target aspect, amplitude scaling, and time-shift misalignment. This paper proposes a hybrid deep learning architecture—integrating 1D Convolutional Neural Networks (CNNs), Bidirectional Recurrent Neural Networks (BiRNNs), and an attention mechanism—to address these challenges. By combining the spatial feature extraction of CNNs with the sequential modeling and "soft alignment" capabilities of BiRNNs and attention, the proposed model achieves a 92.66% recognition rate and demonstrates remarkable robustness to target misalignment compared to traditional CNN-based approaches.

Introduction

In the field of Radar Automatic Target Recognition (RATR), a High-Resolution Range Profile (HRRP) serves as a 1D signature representing the distribution of a target's scatterers along the radar line-of-sight. While HRRPs are rich in structural information and computationally efficient, they suffer from significant sensitivities, most notably the "time-shift" or misalignment problem. If a target is not perfectly centered in the radar's observation window, traditional models—particularly Convolutional Neural Networks (CNNs)—often fail because their learned features are tied to specific spatial locations.

Cessna Citation S/II aircraft and its corresponding High-Resolution Range Profile (HRRP) showing normalized amplitude across range cells.

Methods

The researchers introduce a three-part pipeline designed to be trained end-to-end:

  1. 1D-CNN Layers: The base of the architecture uses convolutional layers to explore spatial correlations in raw HRRP data, extracting expressive features that are more robust than raw data points.
  2. Bidirectional RNN (BiRNN): The features from the CNN are treated as a sequence. The BiRNN processes this sequence in both forward and backward directions, capturing the temporal dependence between different range cells to understand the global structure of the target.
  3. Attention Mechanism: An attention module calculates weights for each hidden state, allowing the model to automatically focus on the most discriminative parts of the signal (the actual target) while ignoring noise or empty range cells. This "soft alignment" is the key to resisting time-shifts without tedious manual preprocessing.

Key Results and Insights

The study utilized real-world measured HRRP data from three aircraft: the An-26, Cessna Citation S/II, and Yark-42. The proposed CNN-BiRNN-3 model achieved an average recognition rate of 92.66% on aligned data, outperforming standard CNNs (92.57%) and the previous state-of-the-art TARAN model (90.10%).

Crucially, the hybrid model demonstrated superior robustness to misalignment. In tests shifting the target by ±16 range cells, the hybrid model's accuracy remained nearly constant, whereas the standard CNN's performance dropped significantly as the shift increased.

Comparison of recognition rates between CNN and the proposed RNN-based hybrid model under varying translation (time-shift) units.

An important discovery was that "deeper is not always better." While more CNN layers increase expressiveness, they also enlarge the receptive field. If the receptive field becomes too large relative to the target, the sequential representation becomes ambiguous. The authors found that a 3-layer CNN provided the optimal balance for sequential modeling.

Conclusion

By combining CNNs, BiRNNs, and attention mechanisms, this architecture provides a robust solution for HRRP-based target recognition that is invariant to time-shifts. While the model excels in high-SNR conditions, future work will focus on improving performance at lower Signal-to-Noise Ratios and scaling the system to handle larger libraries of target types.

Summary updated

Abstract

In this paper, we integrate the advantages of convolutional neural network (CNN) and bidirectional recurrent neural network (BiRNN) with attention mechanism, and propose a CNN-BiRNN based method to recognize the individual high resolution range profile (HRRP). In the proposed method, the CNN is utilized to explore the spatial correlation of raw HRRP data and extract expressive features followed by a BiRNN taking the full consideration of temporal dependence between range cells. Furthermore, in order to enhance the robustness to misalignment, an attentional mechanism is employed after BiRNN to allow the CNN-BiRNN model to focus on the discriminative target area. The combination of CNN and BiRNN with attention mechanism makes the extracted features are not only efficient, but also strongly resistant to the time-shift sensitivity. Experimental results on measured HRRP data demonstrate the effectiveness and the robustness to misalignment of the proposed method.
BibTeX
@article{wan2020recognizing,
  author = {Jinwei Wan and Bo Chen and Yingqi Liu and Yijun Yuan and Hongwei Liu and Lin Jin},
  title = {{Recognizing the HRRP by Combining CNN and BiRNN With Attention Mechanism}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2969450},
}

Loading related content...

Figure