Expiry date extraction from pharmaceutical products using deep learning techniques for visually impaired individuals

Accurate extraction of expiry dates from pharmaceutical packaging is essential to ensure consumer safety, regulatory compliance, and efficient machine-based inventory tracking. Yet, automated recognition is threatened by font variations, text direction, background noise, and image distortion. This paper introduces a comparison between deep learning-based and conventional Optical Character Recognition (OCR) methods for extracting expiry dates. We compare ResNet-50 for feature extraction, Fully Convolutional Networks (FCN) for date region segmentation, Faster R-CNN for text detection, and Tesseract OCR, EasyOCR, TrOCR, and CRNN for text recognition. Four heterogeneous datasets of medicinal product images are utilized to compare detection accuracy, recognition precision, robustness to noise, and computational cost. Though light-weighted OCR approaches such as Tesseract and EasyOCR are simple, they perform poorly with low contrast and skewed text. FCN and Faster R-CNN enhance text localization, while TrOCR and CRNN improve recognition using deep sequence modeling. Experimental comparisons account for inference time, parameter efficiency, and loss functions like Huber Loss, Mean Squared Error, and Cross-Entropy. Our results reveal trade-offs between accuracy and computational cost, providing insights for selecting appropriate models based on real-world constraints. This work supports the development of scalable, automated expiry date detection systems, thereby enhancing their reliability in pharmaceutical and retail contexts.