Figure 1

The overall architecture of the proposed method (the input image is processed through the visual module, where ResNet-45 extracts features related to the structure and strokes of Yi characters, while Deformable DETR captures spatial relationships and deformations, effectively addressing the distinctive characteristics of Yi characters. These features are passed through a Linear layer and a softmax layer to generate the visual prediction map, which is then input into the linguistic model. The Pyramid Pooling Transformer reduces sequence length and enhances feature representation. The linguistic features are then processed through feed-forward layers to generate the linguistic prediction map. Although the linguistic prediction map does not contribute to the final output, it is crucial for training the linguistic model, improving its ability to extract accurate features for Yi character recognition. Finally, Cross Attention aligns the visual and linguistic features, and linear and softmax operations generate the final prediction result).