Differentiable Architecture Search with Multi-Domain Time-Frequency Features for Underwater Target Recognition

Underwater target recognition using active sonar echoes has gained significant research interest due to its critical role in marine surveillance and security. However, the sonar sensors encounter disturbances from seafloor reverberation and the complicated noise background. Traditional classification methods rely heavily on manual feature extraction and expert-defined systems, hindering their adaptability and robustness. In this paper, we propose a novel differentiable architecture search algorithm (DARTs-MTF) with multi-domain time-frequency features for underwater target recognition. Initially, we apply three different time-frequency transformation methods to the preprocessed active sonar echoes to extract robust echo features. We obtain the comprehensive feature dataset through the encoder-decoder network. By redesigning the search space of differentiable architecture search, the optimal network architecture is automatically identified from the extracted feature dataset. Extensive experiments conducted on South China Sea trial datasets demonstrate that the proposed DARTs-MTF approach achieves superior performance compared to existing manually designed classifiers and neural architecture search methods, achieving an overall accuracy of 83.08%.