Feature extraction of passive sonar target based on two cepstrums

Aim. The classification experience of proficient sonar operators is valuable. We try to embody as much as possible such experience in our proposed classifier. In the full paper, we explain in some detail our method and its effectiveness. In this abstract, we just add some pertinent remarks to listing the two topics of explanation. The first topic is: the description of the cepstrums of noise signals. Its four subtopics are: cepstrums (subtopic 1.1), the properties of cepstrums (subtopic 1.2), linear prediction coefficient (LPC) cepstrum (subtopic 1.3) and the Mel cepstrum (subtopic 1.4). The second topic is: the extraction and classification of the features of the two cepstrums. Its three subtopics are: feature extraction and performance analysis (subtopic 2.1), the design of a classifier (subtopic 2.2) and classification experiments and their results (subtopic 2.3). In subtopic 2.1, we use the LPC cepstrum and the Mel cepstrum to obtain from target-radiated noise signals the impulse response of a sounder in the cepstrum domain and extract the feature vector of the impulse response. In subtopic 2.2, we design a BP (Back Propagation) neural network classifier. In subtopic 2.3, we did classification experiments on three classes of passive sonar targets. The analysis of experimental results presented in Tables 3 and 4 in the full paper, shows preliminarily that the total recognition rates of the classifier are 83.9% and 84.2% respectively for the LPC cepstrum and the Mel cepstrum.