Abstract: Aimed at applying the sound recognition technology on the target recognition in border and coastal surveillance systems

a new method of feature extraction was proposed based on the optimum wavelet packet basis. According to the frequency characteristics of the collected border and coastal surveillance sound signals

these sound signals were analyzed by selective multi-scale wavelet packet decomposition and the optimum wavelet tree was obtained. From the high dimensional coefficients of the corresponding optimum wavelet packet basis

the meaningful and compact energy feature vectors were built. After that

these vectors were inputted in a BP neural network to train the network. The trained BP network was then used to detect border and coastal surveillance sound signals. The simulation results indicated that the identification ratio is up to 94% using this feature extraction method under noisy environment

improved 6% compared with the method based on wavelet packet decomposition.