This paper proposes a convolution-based mixture of experts (CMoE) approach for underwater acoustic target recognition, addressing the challenges of high intra-class diversity and inter-class similarity in underwater acoustic signals. The CMoE model consists of multiple expert layers and a routing layer that adaptively assigns inputs to the most suitable expert based on high-level representations. This design enables the model to learn complex underwater signals with multiple independent parameter spaces. Additionally, the model incorporates balancing regularization and an optional residual module to optimize performance. The proposed method is validated on three underwater acoustic databases, demonstrating significant improvements in recognition accuracy compared to existing methods. The results show that CMoE effectively captures latent characteristics from high-level representations and adapts to diverse underwater data. Visualization analysis further confirms the model's ability to handle inter-class similarity and intra-class diversity. The study highlights the importance of balancing regularization in preventing overfitting and improving model generalization. While the CMoE approach shows promising results, further research is needed to fully exploit the potential of the mixture of experts paradigm in underwater acoustic recognition. The work contributes to the development of robust and accurate underwater acoustic target recognition systems.This paper proposes a convolution-based mixture of experts (CMoE) approach for underwater acoustic target recognition, addressing the challenges of high intra-class diversity and inter-class similarity in underwater acoustic signals. The CMoE model consists of multiple expert layers and a routing layer that adaptively assigns inputs to the most suitable expert based on high-level representations. This design enables the model to learn complex underwater signals with multiple independent parameter spaces. Additionally, the model incorporates balancing regularization and an optional residual module to optimize performance. The proposed method is validated on three underwater acoustic databases, demonstrating significant improvements in recognition accuracy compared to existing methods. The results show that CMoE effectively captures latent characteristics from high-level representations and adapts to diverse underwater data. Visualization analysis further confirms the model's ability to handle inter-class similarity and intra-class diversity. The study highlights the importance of balancing regularization in preventing overfitting and improving model generalization. While the CMoE approach shows promising results, further research is needed to fully exploit the potential of the mixture of experts paradigm in underwater acoustic recognition. The work contributes to the development of robust and accurate underwater acoustic target recognition systems.