Underwater target recognition method based on multi-domain active sonar echo images
- Resource Type
- Conference
- Authors
- Wang, Qingcui; Du, Shuanping; Wang, Fangyong; Chen, Yuechao
- Source
- 2021 IEEE International Conference on Signal Processing, Communications and Computing (ICSPCC) Signal Processing, Communications and Computing (ICSPCC), 2021 IEEE International Conference on. :1-5 Aug, 2021
- Subject
- Communication, Networking and Broadcast Technologies
Computing and Processing
Signal Processing and Analysis
Deep learning
Training
Image recognition
Target recognition
Oceans
Sonar
Signal processing
Underwater target recognition
deep learning
active sonar echo images
multi-domain
- Language
The classification and recognition of underwater target by active sonar echo remains a challenging task due to the complex ocean environment and the multiple interferers in the sea. In this paper, an underwater target recognition method is proposed based on multi-domain active sonar echo images. The active sonar echo is first preprocessed to generate images in multiple domains. Then a deep neural network is constructed which is composed of a shared network and several domain-specific attention modules. The shared network is trained on images in all domains to get the global generalized features. The domain-specific features are then further extracted from the global feature through the attention module in each domain. The co-utilization of images in all domains enlarges the data size for training and enhances the feature representation ability of the model. Experiment results demonstrate that the features extracted from the proposed method get better recognition performance than network trained on images in single domain.