Mel Spectrum Feature Recognition Based on Adaptive Center Frequency

Abstract

The key to underwater acoustic target recognition is to extract the line spectrum feature of the ship target. The perception of auditory features by the Mel spectrum is similar to the human ear, and the expression is apparent in the low-frequency bands, suitable for feature extraction. However, the traditional Mel filter has fixed structural parameters and is not sufficiently associated with the sample. Based on this, we propose an adaptive Mel spectrum generation method based on deep learning methods. The relationship between the center frequency of the Mel filter bank and the sample data is established using the data-driven approach based on the computing power of the neural network. In order to verify the effectiveness of this method, comparative experiments were carried out in the final part. The results showed that compared with the traditional Mel spectrum, the accuracy of the adaptive Mel spectrum proposed in this paper was increased by 4.2%, which verified its practicability and feasibility in feature extraction.