A Lightweight UAV Audio Detection Model Based on Multiscale Feature Fusion

• Published in: 2023 5th International Academic Exchange Conference on Science and Technology Innovation (IAECST) pp. 1524 - 1528
• Main Authors: Li, Tao, Huang, Zhihua, Zhai, Xianxu, Wang, Siyuan
• Format: Conference Proceeding
• Language: English
• Published: IEEE 08.12.2023
• Subjects: Autonomous aerial vehicles; Computational modeling; Data models; Deep Learning; Lightweight network; Mel-spectrogram; Performance evaluation; Real-time UAV detection; Security; Technological innovation; Training; UAV sound detection
• DOI: 10.1109/IAECST60924.2023.10503266

With the widespread use of drones, drone audio detection algorithms have gained more and more attention in recent years. By analyzing the sound emitted by drones, automatic detection and fast identification of drones can be realized to avoid the security threats brought by drones. However, the existing drone audio detection algorithms have the problems of a single data type in the training dataset and a large and complex detection model, which leads to inaccurate detection of the model in practical applications or needs to be more Lightweight to deploy on embedded devices. To address the above problems, this paper first constructs a large-scale UAV audio dataset containing hovering and flight states and proposes a lightweight audio detection algorithm. Based on the MobileNetV3-Small model, a multi-scale feature fusion structure and an efficient multi-scale attention module are added to improve the expression ability of features; the Bneck module is designed by combining the ECA attention mechanism and ordinary convolution, and the network is also pruned to reduce the complexity of the model. Experimental results on self-built and open datasets show that our model obtains high-performance detection results with a lightweight network architecture. On the self-built dataset, the detection accuracy reached 95.73%, the detection precision reached 98.04%; on the public dataset, the accuracy was improved to 97.88%, the precision was improved to 99.83%. However, the trainable parameter size of the model was only 0.84M, and the overall model size was only 3.30M.