Correction of OCR results using large language models
Zhouzhen Shi, Yi Chen · 2025
This paper proposes a large language model (LLM)-based method for correcting OCR results, integrating prompt engineering with recognition models including PP-OCRv4 and GLM-4-Flash. By introducing an LLM-driven error correction mechanism, the accuracy of OCR-recognized texts is significantly improved, particularly in punctuation restoration and character rectification. Experimental results demonstrate that for datasets with GLM correction confidence scores below 0.95, the proposed method achieves an accuracy improvement of approximately 17%. Specifically, the GLM-enhanced framework outperforms standalone PP-OCRv4 by 4.7% and surpasses the Kimi large model by 29.5% in accuracy. These findings validate the high practical applicability of the GLM-based correction strategy in real-world scenarios.