SVR-based approach to improve active sonar detection in reverberation

Whitening method is widely used for improving active sonar detection in reverberation environment, which is equivalent to AR model estimation. However, traditional whitening methods suffer from several problems due to the varying statistics and nonlinearity of reverberation noise. In this paper, we use Support Vector Regression (SVR) to obtain the parameters of a whitening filter. The algorithm of SMO without bias is used to train SVR and three speed-up approaches are proposed. The SVR parameters C and p are selected by evaluating the detection performance. The ability of SVR prewhitener is verified on real lake data. Experimental results show that SVR prewhitener outperforms traditional methods significantly and provides an excellent performance even under low signal-to-reverberation ratio (SRR) and low-doppler conditions.