A Robust Framework for Coffee Bean Package Label Recognition: Integrating Image Enhancement with Vision-Language OCR...

Area of Science:

• Computer Vision
• Natural Language Processing
• Machine Learning

Background:

• Text recognition on coffee bean packaging is crucial for product tracking and brand verification.
• Challenges include variable image quality, diverse packaging, and environmental factors.

Purpose of the Study:

• To develop an advanced Optical Character Recognition (OCR) pipeline for coffee bean package labels.
• To improve text recognition accuracy and reliability in real-world scenarios.

Main Methods:

• A pipeline combining image enhancement techniques and a Vision-Language (VL) Qwen VL-based OCR model with structured prompts.
• Development and evaluation of four Qwen-VL OCR variants using prompt engineering.
• Creation of a coffee bean package image dataset (LRCB and HRCB) with diverse real-world challenges.
• Comparison against baseline OCR models: DocTR, PaddleOCR, EasyOCR, and Tesseract.

Main Results:

• The proposed Qwen-VL OCR variants demonstrated significant improvements over baseline methods.
• Evaluation metrics included Levenshtein distance, Cosine similarity, Jaccard index, Exact Match, BLEU, and ROUGE scores.
• The framework showed strong generalization capabilities on the public POIE dataset.

Conclusions:

• The developed pipeline offers a practical and reliable solution for coffee bean package label recognition.
• The approach effectively addresses challenges posed by image quality and environmental variations.
• This work advances the state-of-the-art in specialized OCR applications.