Studies On Physiological Perception Feature Extraction Methods In Underwater Target-radiated Noise

Underwater target feature extraction method is one of the key obstacles to sonartarget recognition system for further development. Since that underwater targets arechanging rapidly and sensitive to conditions, speed, hull structure and location ofcollection in a complex ocean, the performance of sonar recognition system degradesrapidly when using the traditional feature extraction methods. Therefore, in order toimprove the robustness of sonar recognition system, this thesis took human auditorycharacteristics into account, and applied them to characterize underwater targets.The main contributions are as follows:1. Some commonly used pattern recognition methods are compared. BP neuralnetwork and support vector machine (SVM) are selected to recognize and classify thetargets for their superiority in processing non-linear and small sample of underwatertarget-radiated noise. Experiment results show that the recognition algorithm of alltesting samples by using BP neural network and SVM reaches satisfactory accuracy,which suggest that they are the effective methods to recognize and classify thefeatures of underwater target-radiated noise.2. Psychoacoustic parameters are applied to reflect differences in auditoryperception quantitatively since they can reflect people’s subjective feelings. Thispaper describes the calculation processing of loudness features and improves thealgorithm. Experiment results show that the modified algorithm is effective andpracticable, which can avoid calculating the incentive of signal and easily extract theloudness feature in the frequency domain. Moreover, the modified algorithmprovides an effective computational tool for the further study of auditory features.3. Through analyzing and comparing with the performance of different auditoryfilters, some feature extraction techniques based on auditory characteristics areproposed: gammatone frequency discrete wavelet coefficients, auditory cepstralcoefficients, auditory slow feature analysis, gammachirp cepstral coefficients, andmodified cochlear filter analysis model. By using the mathematical model ofauditory system, these methods can accord well with the treatment process of ear,extract underwater target-radiated noise feature, avoid losing the key information andimprove the anti-interference performance as well as the recognition accuracies of thesystem. 4. Conventional spectral compression scheme would over-compress somefrequency components and under-compress other frequency components of signal atthe same time. To solve this problem, we proposed feature extraction techniquebased on perceptual non-uniform spectral compress (PNUSC) according to the powerlaw of hearing and auditory masking. Applying PNUSC technique can effectivelyimprove the accuracies of the underwater targets recognition. Dealing with thedecline of recognition rate of sonar system in the noisy environments, the improvedperceptual non-uniform spectral compress (IPNUSC) feature extraction algorithm isput forward. Through experiments on the three classes of targets, the results showthat the IPNUSC algorithm can effectively enhance the robustness and improveaccuracies of underwater target recognition.5. Inspired by the partial masking of background noise on a signal, an SNR(signal-to-noise ration)-dependent non-uniform spectral compression (SDNUSC)method is proposed to better simulate the intensity-to-loudness conversion. Besides,parameter selection is analyzed and given in this thesis. Recognition results showthat the SDNUSC front-end can deal with different types of additive noises withperformance significantly better than that of MFCC, LPCC, PLP front-end and otherPNUSC scheme. In addition, the feature vectors extracted are more robust.6. A feature extraction technique based on broadband constant beamwidthbeamforming and auditory perception is proposed to improve recognition capabilityof sonar. To adapt to the complexity of battlefield and demands of variousinformation of targets, as well as improving recognition capabilities of Sonar systems,no distortion is expected at the output of wideband receiving array. Therefore,constant width beamformer is designed and applied for classification experimentusing simulated array data. Experiment results show that the proposed methodproduces higher recognition accuracies than other feature extraction methods using asingle hydrophone. Meanwhile, the proposed method has good tolerance for thestate of underwater targets.