A transfer learning method using speech data as the source domain for micro-Doppler classification tasks

摘要

In recent years, micro-Doppler target classification technology has been widely used for radar target recognition. However, due to the lack of sufficient data, it has become a challenge to train a model with excellent performance using the transfer learning method. Most of the existing transfer learning methods for micro-Doppler tasks use optical image data or simulation data as the source domain, and the use of fine-tuning as the transfer method makes it difficult to obtain good results. This paper proposes a transfer learning method using speech data as the source domain for micro-Doppler classification tasks. The proposed method uses speech data as the source domain and improves the accuracy of micro-Doppler classification through TCA and deep learning models used jointly. After experimental verification, the proposed method can use the 2.8 M parameters to improve accuracy by more than 5% compared with common methods in the case of a small number of frames, and the proposed method achieves better results with a small number of points.