一种针对甲骨文单字图像推理任务的训练方法和装置
By optimizing the training scheme through multi-task collaboration and using a multimodal large model to train the structure of oracle bone characters in stages, the problem of image recognition and interpretation of single oracle bone characters was solved, and accurate interpretation and interpretable output of complex oracle bone characters were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2025-07-01
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies struggle to effectively recognize and interpret single-character images of oracle bone inscriptions, especially in complex character formation logic and semantic reasoning. General multimodal large models lack applicability, and existing methods rely on limited labeled data, making it difficult to interpret unknown oracle bone inscriptions.
A multimodal large model and multi-task collaborative optimization training scheme is adopted. The backbone network is trained to analyze the shape structure of oracle bone characters. The training is carried out in stages for Chinese character recognition, modern Chinese character radical decomposition, character period identification, oracle bone image recognition and character shape thinking chain reasoning. The model parameters are fine-tuned by combining the reward function to generate structured output.
It significantly improves the accuracy and interpretability of the model's interpretation of oracle bone script, providing a structured, low-cost, automated solution capable of handling the understanding and interpretation of complex oracle bone script characters.
Smart Images
Figure CN120997850B_ABST