Existing vision-based maritime target recognition methods suffer significant performance degradation under conditions of low-resolution cameras or occlusion. The navigational data in AIS signals can provide supplementary information to alleviate this issue. However, constructing a robust maritime target recognition model using multimodal sensor data remains a challenge due to the differences between modalities. This paper proposes a dual-stream heterogeneous collaborative learning method based on DeepSeek, which guides maritime target recognition by extracting information from visual sensors. It also improves model robustness by utilizing feature distillation to reduce the impact of noise. To achieve model lightweighting, knowledge from DeepSeek is transferred to a student network, significantly reducing the model’s parameter size and computational complexity while maintaining performance. The F1 score and accuracy on a self-constructed maritime multimodal dataset demonstrate the superior performance of the DeepSeaNet method compared to existing state-of-the-art maritime target recognition methods.