The Transformer shows good prospects in computer vision. However, the Swin Transformer model has the disadvantage of a large number of parameters and high computational effort. To effectively solve these problems of the model, a simplified Swin Transformer (S-Swin Transformer) model was proposed in this article for handwritten Chinese character recognition. The model simplifies the initial four hierarchical stages into three hierarchical stages. In addition, the new model increases the size of the window in the window attention; the number of patches in the window is larger; and the perceptual field of the window is increased. As the network model deepens, the size of patches becomes larger, and the perceived range of each patch increases. Meanwhile, the purpose of shifting the window’s attention is to enhance the information interaction between the window and the window. Experimental results show that the verification accuracy improves slightly as the window becomes larger. The best validation accuracy of the simplified Swin Transformer model on the dataset reached 95.70%. The number of parameters is only 8.69 million, and FLOPs are 2.90G, which greatly reduces the number of parameters and computation of the model and proves the correctness and validity of the proposed model.

Recently, Vision Transformer (ViT) has been widely used in the field of image recognition. Unfortunately, the ViT model repeatedly stacks 12-layer encoders, resulting in a large number of model computations, many parameters, and slow training speed, making it difficult to deploy on mobile devices. In order to reduce the computational complexity of the model and improve the training speed, a parallel and fast Vision Transformer method for offline handwritten Chinese character recognition is proposed. The method adds parallel branches of the encoder module to the structure of the Vision Transformer model. Parallel modes include two-way parallel, four-way parallel, and seven-way parallel. The original picture is fed to the encoder module after flattening and linear embedding processing operations. The core step in the encoder is the multihead attention mechanism. Multihead self-attention can learn the interdependence between image sequence blocks. In addition, the use of data expansion strategies increases the diversity of data. In the two-way parallel experiment, when the model is 98.1% accurate on the dataset, the number of parameters and the number of FLOPs are 43.11 million and 4.32 G, respectively. Compared with the ViT model, whose parameters and FLOPs are 86 million and 16.8 G, respectively, the two-way parallel model has a 50.1% decrease in parameters and a 34.6% decrease in FLOPs. This method has been demonstrated to effectively reduce the computational complexity of the model while indirectly improving image recognition speed.