Scholarlink

This paper proposes a hybrid solution to extract expiration dates from real packaged product images using three OCR modules. We assess Tesseract OCR, EasyOCR and PaddleOCR, using a dataset of 665 sample images. These OCRs are in conjunction with the Llama 3.2 3B model, used exclusively as a deterministic post-processing filter that selects the most plausible expiration date from the OCR outputs to improve reliability. This study examines how OCR models behave when filtered by an LLM selector, using match, unknown, and mismatch distributions, as well as similarity cluster analyses to assess prediction variability. The findings reveal strong performance discrepancies between individual modules. PaddleOCR shows the highest accuracy, followed by EasyOCR and Tesseract with the lowest.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

OCR and LLM Pipeline for Reliable Expiration Date Reading in Accessibility Applications

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Theodor-Radu Grumeza,
  • Alexandra-Emilia Fortiș

摘要

This paper proposes a hybrid solution to extract expiration dates from real packaged product images using three OCR modules. We assess Tesseract OCR, EasyOCR and PaddleOCR, using a dataset of 665 sample images. These OCRs are in conjunction with the Llama 3.2 3B model, used exclusively as a deterministic post-processing filter that selects the most plausible expiration date from the OCR outputs to improve reliability. This study examines how OCR models behave when filtered by an LLM selector, using match, unknown, and mismatch distributions, as well as similarity cluster analyses to assess prediction variability. The findings reveal strong performance discrepancies between individual modules. PaddleOCR shows the highest accuracy, followed by EasyOCR and Tesseract with the lowest.