Feature Analysis of Passive Underwater Targets Recognition Based on Deep Neural Network

As an important research direction of sonar signal processing, the research on taking deep neural network (DNN) as a classifier for underwater target recognition has become a hotspot. Feature extraction methods are adapted to the classifier and WPCE, FBANK and MFCC features was normally used as input for DNN. In this paper, a time-delayed neural network (TDNN) based classification system has been proposed to compare the recognition effects of different features. The wavelet packet component spectrum (WPCS) was proposed as a new feature extraction method based on WPCE, since the wave packet transform separating noise for line spectrum extraction. The experiments on the real recorded data of two type of vessels indicate that the proposed system using WPCS features provides higher recognition accuracy than using WPCE, FBANK and MFCC features. Feature vector extension and average were used to optimize WPCE features which achieved 2.93% and 9.76% improvement respectively.