Knowledge-Enhanced TrOCR for Automated Storage and Retrieval Systems: A Novel Approach to Pharmaceutical Package Text Recognition

Accurate cargo information acquisition is crucial for ensuring inventory precision and operational efficiency in automated storage and retrieval systems (AS/RS). Traditional methods including manual input, barcode scanning, and conventional optical character recognition (OCR) face significant limitations in complex warehouse environments with varying lighting conditions, text distortions, and domain-specific terminology. This paper proposes KAT-OCR (KnowledgeAugmented TrOCR), a novel text recognition framework that integrates domain knowledge with transformer-based OCR to address the challenges of pharmaceutical package text recognition. Our approach consists of two key innovations: the KnowledgeEnhanced Vision Encoder (KEVE) that dynamically fuses visual features with retrieved domain knowledge, and the KnowledgeConstrained Text Decoder (KCTD) that leverages domain constraints during text generation. Experimental results on the GY-PackInfo dataset demonstrate that KAT-OCR achieves 96.3% character accuracy and 81.7% key information accuracy, representing improvements of 3.2% and 9.2% respectively over vanilla TrOCR, while maintaining real-time inference speed (92 ms per image). The proposed method shows particular robustness in challenging scenarios including low-light conditions (+9.3% improvement), partial occlusion (+16.5% improvement), and confusing character recognition (+ 1 8. 4% improvement).