For identifying each vessel from ship-radiated noises with only a very limited number of data samples available, an approach based on the contrastive learning was proposed. The input was sample pairs in the training, and the parameters of the models were optimized by maximizing the similarity of sample pairs from the same vessel and minimizing that from different vessels. In practical inference, the method calculated the distance between the features of testing samples and those of registration templates and assigned the testing sample into the closest templates for it to achieve the parameter-free classification. Experimental results on different sea-trial data demonstrated the advantages of the proposed method. On the five-ship identification task based on the open-source data, the proposed method achieved an accuracy of 0.68 when only five samples per vessel were available, that was significantly higher than conventional solutions with accuracies of 0.26 and 0.48. Furthermore, the convergence of the method and the behavior of its performance with increasing data samples available for the training were discussed empirically.