Deep Learning-Driven Expiry Date Recognition on Medicine Bottles via YOLOv8 Segmentation and Multi-Stage Image Denoising

Automated expiry date recognition (EDR) on pharmaceutical packaging is essential for ensuring medicine safety and minimizing waste, but it poses challenges due to text unpredictability, environmental interference, and intricate label geometries. This study presents a comprehensive deep learning system that integrates sophisticated picture pre-processing with YOLOv8-based instance segmentation to overcome these restrictions. A curated dataset including 1,000 high-resolution photos of pharmaceutical bottles, encompassing various lighting situations, camera angles, and date formats, was assembled. The pre-processing pipeline incorporates wavelet denoising, BM3D filtering, and contrast-limited adaptive histogram equalization (CLAHE) to alleviate glare and enhance low-contrast text artefacts. The advanced YOLOv8 architecture utilizes multi-scale feature fusion for accurate text localization on curved and uneven surfaces. Comparative assessments reveal the framework's superiority over leading models (Mask R-CNN, U-Net, and FCN) in segmentation precision, attaining a 95.7% F1 score and a 34% decrease in boundary error (ASD). Ablation research verifies the impact of each pre-processing step. The technology, in conjunction with an OCR module, facilitates comprehensive expiry date extraction with a character error rate (CER) of 0.9% under optimum settings. The method, although based on a restricted dataset, demonstrates significant potential for real-time quality management in pharmaceutical supply chains, enhancing AI-driven compliance monitoring and sustainable healthcare practices.