The sample number of ship target categories collected in actual scenes is not balanced, and the model training easily leads to be overfitting. The data set of the traditional transfer learning is divided into categories, which results in low recognition accuracy of unlabeled new categories. To solve the above problems, a small sample ship identification algorithm based on the fusion of the crosstarget universal global attention mechanism and the relationship measurement network is proposed. This method introduces the universal attention mechanism into the relation network, uses the original features extracted by relation network, and smooths the target features between imbalanced categories through the universal attention mechanism, and compares them with the original features extracted by the relation network. After feature fusion, feature distance measurement is performed. This method enhances the consistency among universal features, which is conducive to learning invariant target features and improve the performance of ship recognition with few samples and few labels. In this way, the overfitting problem caused by imbalance of categories in the training process could be solved. Using the ship data set collected and produced by ourselves to test the proposed method, the recognition accuracy is improved 5. 6% (5- shot) and 3. 2% (1-shot). The impact of imbalanced category on the model ship recognition is reduced, and the robust of the model is enhanced.