A Joint Framework with Audio Generation for Rare Gunshot Event Detection

Nowadays, gunshot detection is closely related to social security, but it needs more attention. The data driven method for gunshot detection which a large corpus of gunshots will be needed for training a neural network is urgently needed. To address this requirement, we propose a novel unified framework, the Gunshot Generation and Detection (GGD), specifically designed for gunshot event detection. It merges an audio generation model with a detection model to partially alleviate the issue of data scarcity problem. Comparative analysis indicates that the GGD model surpasses non-generative models. Remarkably, it outperforms models employing data augmentation techniques. Furthermore, the GGD framework is easy to incorporate with diverse detection network architectures, such as VGGish and Mobile Net. When coupled with a Convolutional Neural Network (CNN), our methodology yields recall varying from 93.98% to 98.20%. These findings demonstrate that this integrated approach significantly enhances the detection performance of gunshot detection models.