Abstract: In order to address the interference caused by the time-varying characteristics of underwater acoustic channels in non-cooperative underwater acoustic communication signal recognition and to meet the needs of low-power deployment, deep learning recognition requires lightweight design to improve recognition accuracy and enhance model generalization ability. A lightweight and efficient recognition model is proposed firstly based on an improved DenseNet structure. By adopting dimension transformation and model compression methods, the model structure and parameters are optimized, reducing the complexity of model inference while ensuring recognition accuracy. Secondly, a multi-modal expression fusion strategy is employed, effectively combining features extracted by different networks, fully utilizing the complementarity of information, thereby significantly improving recognition accuracy. On the simulated dataset, the fusion network achieves a recognition rate of over 94.65% at a signal-to-noise ratio of −6 dB and 98.03% at 0 dB. On the real measured dataset, the accuracy of the base network after transfer learning reaches 98.05%. Lake test results validate the effectiveness of the proposed method.