Related Experiment Videos
A lightweight hybrid deep learning framework for multi-pill detection, multi-attribute recognition, OCR-based imprint
1School of Computer Science Engineering and Information Systems, Vellore Institute of Technology, Vellore, Tamil Nadu, India.
Frontiers in Artificial Intelligence
|August 4, 2026
Summary
This study introduces a lightweight AI framework for accurate pill recognition, integrating detection, classification, and imprint analysis to reduce medication errors and enhance patient safety.
Area of Science:
- Artificial Intelligence
- Computer Vision
- Pharmacovigilance
Background:
- Adverse drug events (ADEs) are a significant cause of preventable harm, often stemming from medication identification errors.
- Existing automated pill recognition systems often treat detection, classification, and imprint recognition as separate tasks, limiting real-time healthcare application.
- Older adults, visually impaired individuals, and those with limited health literacy are particularly vulnerable to medication errors.
Purpose of the Study:
- To develop a lightweight, unified AI framework for multi-pill and multi-attribute recognition.
- To support pharmacovigilance and AI-driven clinical decision support systems.
- To create a system suitable for real-time healthcare deployment and medication identification.
Main Methods:
- A hybrid framework integrating YOLOv2n for pill detection, MobileNetV4-ConvSmall for classification, and EasyOCR for imprint recognition.
- Development of a custom dataset with 400 pill classes (20,000 images) for classification and 40,000 labeled pill instances for detection.
- Performance evaluation through comparative experiments, functional comparisons with existing systems, and cross-device qualitative assessments.
Main Results:
- The YOLOv2n detector achieved high precision (0.963) and recall (0.982).
- MobileNetV4-ConvSmall demonstrated excellent classification accuracy (99.60%), excelling in shape and color recognition.
- The integrated framework achieved Top-1 accuracy of 89.83% and Top-5 accuracy of 99.49%, with robust performance across devices and varying image conditions.
Conclusions:
- Lightweight deep learning architectures can enable scalable medication identification solutions.
- The unified framework enhances medication identification and supports assistive healthcare workflows.
- Clinical judgment and medication verification must remain under healthcare professional supervision.