Active sonar target recognition based on sub-beam filling and time-frequency fusion tensor

This paper addresses the issues of low signal-to-noise ratio (SNR), the loss of spatial information in conventional beamforming, and small sample in underwater active sonar detection and recognition. A feature extraction method based on sub-beam filling (SBF) and time-frequency tensor feature fusion is proposed to enhance the feature extraction capability of underwater sonar echoes, which integrates frequency-weighted time-frequency features from STFT, CWT, and SPWVD. To further improve the recognition performance, an improved ConvNeXt architecture, ConvNeXt-DCA, is proposed, incorporating asymmetric convolution kernel decomposition and a lightweight Channel Aggregation (CA) module. Experimental evaluations on both pool and sea trial datasets demonstrate the superiority of the proposed method. Compared to standard beamforming, SBF improves average accuracy from 67.1% to 80.6%. The ConvNeXt-DCA model achieves the highest recognition accuracy of 92.1% on the sea trial dataset and maintains 87.7% under -10 dB SNR in the pool dataset. These results confirm the effectiveness and robustness of the proposed framework in actual sonar recognition scenarios.