Improving Automatic Identification of Medications in Transparent Packaging by Glare Removal and Color Correction

Abstract

Automated Medication Identification (AMI) systems can significantly streamline the daily tasks of pharmacists. Nonetheless, the image analysis methods utilized by AMI systems often encounter difficulties in real-world settings. For example, glare reflections from transparent packaging and color distortions from different lighting conditions may alter the visual appearances of medication images, thereby degrading the recognition accuracy. This paper proposes an innovative approach to mitigate these issues, incorporating image registration techniques to eliminate glare reflections and correct color discrepancies to improve recognition accuracy. The proposed solutions are integrated with a novel ResDenseNet neural network architecture, which efficiently merges cross-level features via skip connections, harnessing the combined merits of ResNet and DenseNet. Empirical evaluations reveal that this integrated solution significantly elevates the recognition rate from 15.19% to 96.85% for a ResDenseNet model trained on a dataset with limited appearance variations. Furthermore, the ResDenseNet outperforms ResNet and DenseNet by 7.5% and 3.3% in recognition rate, respectively.

Keywords

• Deep learning
• medication identification
• color correction
• glare removal
• object detection
• YOLO