Research On Optimized Feature Extraction Technology Used For Automatic Recognition Of Underwater Targets

Acoustic feature extraction of underwater targets is the key issue in the field of underwater signal processing field,which contributes to modern underwater target recognition technology and has extremely important value in both military and civilian applications.Inspired by the superior perceptual performance of the human auditory system,considering that the current sonar monitor still play an irreplaceable role in modern naval combats,this paper aims to construct reasonable optimal features used for underwater acoustic targets,therefore improve the stability of underwater targets under complex noise conditions.By analysis of the underwater acoustic signal with its spectrum characteristics,this paper expounds the basic principle of underwater target recognition system,then proposes several optimized constructions of feature extraction.Firstly,two feature extraction methods based on Mel frequency cepstral coefficients and Gammatone filterbank cepstral coefficients are studied.Although the recognition rate of the latter shows a certain better under low SNR conditions,it always has a shortage of computational complexity.Aiming at this problem,combined with the study of the cochlear perception characteristics,this paper proposes an improved cepstral coefficients extraction based on normalized Gammatone filterbank achieved in frequency domain.At the same time,through the comparison between Mel frequency cepstral coefficients and Gammatone filterbank cepstral coefficients,the effectiveness of normalized Gammatone filterbank has been verified.Additionally,this paper confirmed that the nonlinear compression plays an important role in auditory feature extraction.Based on these conclusions,as well as the advance on noise suppression in the field of speech recognition,this paper further proposes a feature extraction method named power normalized cepstral coefficients based on power bias subtraction.Since features representing single character would be more sensitive to noise,with consideration that auditory features really ignore the phase information of targets,feature representing instantaneous frequency extracted on the sub-bands can compensate the amplitude features.Based on the amplitude weighted smooth estimation method and optimized subband instantaneous frequency estimation method,combined with the improved GFCC feature as well as robust PNCC feature,this paper constructs four fusion feature extraction methods used for underwater target automatic recognition,fusion feature extraction based on GFCC and subband instantaneous frequency,fusion feature extraction based on GFCC and optimized subband instantaneous frequency,fusion feature extraction based on PNCC and subband instantaneous frequency,and fusion feature extraction based on PNCC and optimized subband instantaneous frequency.In order to implement the effective evaluation of different features,this paper researched two classifier models,BP neural network and support vector machine.As results of the latter one showing better performance and advantage on evaluation ability in underwater recognition domain,this paper adopts Support Vector Machine as the classifier to carry out feature extraction methods proposed in this paper.Experimental results verify the effectiveness of these proposed feature applied on underwater targets,and further show that the proposed improved and fusion feature extraction methods can maintain excellent recognition results under low SNR conditions,which can effectively improve the stability of underwater target recognition system in complex underwater environment.