The oracle bone inscriptions are a testament to China's rich cultural heritage and is one of the earliest known mature writing systems in human history. Due to their complex textual structure and scarce textual materials, the deciphering of oracle bone inscriptions has become a difficult task. In order to solve these problems, we propose an oracle bone inscriptions recognition method based on spatial transformer network and few-shot learning. Firstly, we utilize the spatial transformer network to appropriately transform the image, thereby enhancing the network's robustness. Secondly, we perform channel segmentation and stripe convolution on the data processed by the spatial transformation network. Finally, we capture the interaction information between different dimensions of the input tensor through the dimension-swapping operation to improve the ability to recognize the oracle. The experiments show that the method exhibits a good recognition function compared with other methods. And the accuracy of 1-shot and 5-shot on our homemade dataset reaches 66.42% and 87.61% respectively.