Quadratic time-frequency feature extraction and fusion for ship targets classification

Abstract

Underwater ship targets classification in a passive mode has valuable applications in many fields. In this paper, a novel time-frequency (TF) feature extraction and fusion strategy based on the ship radiated signal's whole quadratic TF representation matrix is presented. Firstly, with the help of the modified B-distribution, the one-dimensional received signal was transformed into a two-dimensional TFR matrix. Secondly, the TFR matrix singular value decomposition (SVD) based features and the TFR matrix image analysis based features were first introduced to represent a ship type. The fuzzy rough feature selection (FRFS) algorithm was then used for selecting the effective and irreplaceable features from each of the feature subsets, and the final classification vector was obtained through a serial fusion strategy. At last, the classification decision is done by a back propagation neural network (BPNN). Experiments using ship sea trial recorded data show that this feature extraction and fusion scheme in underwater target classification is feasible and the recognition rate reaches 90%.