Improving PaddleOCR optical character recognition for food composition labels using hybrid SymSpell and LayoutLM correction

Abstract

A high level of text recognition accuracy is crucial for optical character recognition (OCR) applications, particularly those designed to extract information from food composition labels. However, packaging materials often introduce distortions, reflections, or irregular printing, which lead to recognition errors and word concatenation. To mitigate this issue, this study proposes a novel post-processing system that integrates PaddleOCR with the spelling correction (SymSpell) algorithm and layout language model (LayoutLM)-guided bigram-based word segmentation to enhance OCR accuracy. The implementation begins with capturing a composition label image using a smartphone camera, followed by initial text extraction performed by PaddleOCR. Consequently, SymSpell corrects misspelled tokens through a domain-specific food and beverage lexicon, while LayoutLM leverages spatial and contextual information to separate concatenated words. The performance was evaluated using the character error rate (CER) metric across four packaging surface categories: flat, curved, reflective, and textured. The results demonstrate consistent improvements across all categories, achieving an average CER of 0.0991, with the greatest reduction on flat surfaces (from 0.0885 to 0.0659). The experiments show that the system is robust across label conditions, lightweight for real-time use, and domain-aware. Its model-agnostic design adds flexibility, enabling effective post-OCR correction across different engines and improving readability for real-world food label applications.