Research Article
Small Sample Underwater Acoustic Target Recognition Based on Full Dimensional Dynamic Feature Enhancement Network
@INPROCEEDINGS{10.1007/978-3-031-60347-1_9, author={Jianxun Tang and Zhe Chen and Mingsong Chen and Junyi Wang and Xiaodong Ma}, title={Small Sample Underwater Acoustic Target Recognition Based on Full Dimensional Dynamic Feature Enhancement Network}, proceedings={Mobile Multimedia Communications. 16th EAI International Conference, MobiMedia 2023, Guilin, China, July 22-24, 2023, Proceedings}, proceedings_a={MOBIMEDIA}, year={2024}, month={10}, keywords={Underwater Acoustic Target Recognition Deep Learning ODConv MHSA PolyLoss}, doi={10.1007/978-3-031-60347-1_9} }- Jianxun Tang
Zhe Chen
Mingsong Chen
Junyi Wang
Xiaodong Ma
Year: 2024
Small Sample Underwater Acoustic Target Recognition Based on Full Dimensional Dynamic Feature Enhancement Network
MOBIMEDIA
Springer
DOI: 10.1007/978-3-031-60347-1_9
Abstract
Due to the complex marine background noise, the existing underwater acoustic target recognition methods are mainly based on a single recognition network, which is difficult to fully consider the multi-faceted characteristics of underwater acoustic targets and to filter and identify target features from the underwater acoustic signals mixed with a large amount of marine noise. At the same time, the traditional underwater acoustic target recognition model is mainly a two-stage model, which cannot meet the existing demand for rapid recognition of underwater acoustic targets. In order to solve the above problems, this paper proposes a full-dimensional dynamic feature enhancement network for small sample underwater acoustic target recognition model (CAODCSP-Darknet), which mainly uses ODCf2MHSA to make the model acquire multiple receptive field gradient flow information in full dimension while focusing on the target, and PloyLoss to customize the classification loss function for small sample underwater acoustic dataset to avoid the model the overfitting problem caused by small samples and imbalance of samples in each category. The experimental results show that the proposed model outperforms the other three baseline models by more than 20% and achieves 98.3% accuracy in identifying four classes of underwater acoustic signals on the Deepship underwater acoustic dataset. The performance is better than existing underwater acoustic target recognition models.
