DEEP LEARNING-BASED OCR FRAMEWORK FOR RECEIPTS: PERFORMANCE EVALUATION OF EAST AND CRNN INTEGRATION

Abstrak

Existing OCR systems often struggle with shopping receipts due to irregular layouts, diverse fonts, and image noise. We propose a domain-specific OCR framework that combines the EAST detector for robust text localisation and the CRNN model for sequence-based recognition. Trained on 320 annotated receipts and tested on 84 images, our system achieved 92.6% character-level and 86.4% word-level accuracy, surpassing Tesseract (+15.2%) and standalone CRNN (+9.7%). These results demonstrate the framework’s effectiveness for receipt-specific OCR, supporting applications such as automated expense tracking and financial record digitisation

Identifier

• DOI: https://doi.org/10.33480/jitk.v11i3.7073
• OpenAlex: https://openalex.org/W7133534304