A Gunshot Recognition Method Based on Multi-Scale Spectrum Shift Module

Abstract

In view of the issues such as the larger network model and lower recognition accuracy of the current gunshot recognition networks, a neural network based on a multi-scale spectrum shift module is proposed in this paper to fully mine the relevant information among the gunshot spectrums. This network employs the architecture of a densely connected convolutional network and uses a multi-scale spectrum shift module on the branch to realize the interaction among spectrum information. This spectrum shift replaces the under-sampling operation among the spectrums, realizes the globalized feature extraction of the spectrum, avoids the loss of information during the under-sampling process, and further improves the quality of the spectrum feature map. Experiments were conducted based on the NIJ Grant 2016-DN-BX-0183 gunshot dataset and YouTube dataset on gunshots that have been open to the public, both of whose classification accuracy reached 83.2% and 95.1%, respectively, with the size of the network model being controlled at around 16 MB. The experimental results indicate that, compared with other existing methods for convolutional neural network, the proposed network can mine globalized time-frequency information better and effectively, and has a higher accuracy of gunshot recognition.