Session: 03-04-03: Artificial Intelligent Applications in Manufacturing III

Paper Number: 165713

Advanced Deep Learning for Pharmaceutical Pill Defect Detection 

 

Quality assurance in pharmaceutical manufacturing represents a critical safeguard for patient safety and regulatory compliance. This research introduces an innovative application of Detectron2, an advanced deep learning framework with instance segmentation capabilities, to automate and enhance the detection of pill defects. Unlike traditional quality control methods relying on human visual inspection or conventional computer vision algorithms, this approach offers superior precision in identifying and localizing various defect types including cracks, chips, discoloration, and contaminants.

The study utilized a comprehensive dataset comprising 780 high-resolution pharmaceutical pill images, strategically partitioned into 400 training, 250 validation, and 130 test samples. Defect annotation was conducted using the Makes Sense AI tool, ensuring standardized and precise labeling of varied defect morphologies. The Detectron2 architecture was selected for its exceptional ability to perform instance-level segmentation, enabling the model to precisely delineate defective regions rather than merely classifying pills as defective or non-defective.

Performance evaluation revealed impressive metrics across multiple dimensions. The model demonstrated a 99% confidence level in defect detection while achieving 98.5% overall accuracy and 97.2% mean Average Precision (mAP). Bounding box detection performance yielded an Average Precision (AP) of 41.95, with corresponding AP50 and AP75 values of 68.58 and 44.06 respectively. More significantly, segmentation performance achieved an AP of 45.14, with AP50 of 68.58 and AP75 of 64.64, confirming the model's capability to accurately define defect boundaries.

Beyond detection, this research incorporated quantitative defect assessment by implementing a reference marker methodology. A calibrated square of known dimensions was positioned alongside pills during imaging, establishing a consistent scale for accurate defect size measurement. This quantification capability transforms the system from a simple detection tool into a comprehensive quality assessment platform capable of objective defect severity evaluation.

The implementation of Detectron2 for pharmaceutical quality control presents multiple advantages over traditional inspection methods. The automated approach significantly reduces inspection time while maintaining consistent detection performance across varied pill types and defect categories. By minimizing human subjectivity and error, the system substantially reduces the risk of defective products reaching consumers—a critical concern in pharmaceutical manufacturing.

This research demonstrates that deep learning-based instance segmentation represents a transformative technology for pharmaceutical quality assurance. The high-precision, real-time detection capabilities offered by Detectron2 provide manufacturers with an efficient tool for meeting stringent regulatory requirements while optimizing production throughput. The methodology established in this study creates a foundation for further integration of AI systems in pharmaceutical manufacturing, potentially extending to additional quality parameters and diverse pharmaceutical formulations to create comprehensive, automated quality management systems.

Presenting Author: Sathish Kumar Gurupatham Kennesaw State University

Presenting Author Biography: Moneesh is an undergrad pursuing his bachelors degree in Information Technology at Kennesaw State University.

Authors:

Joshua Daniel kennesaw state university
Moneesh Rajaram Kennesaw State University
Hygreev Manikandan Kennesaw State University
Sathish Kumar Gurupatham Kennesaw State University