Published Research Paper

This research explores how convolutional neural networks (CNNs) can improve pill image recognition for medical retrieval systems, helping patients and practitioners identify unidentified medications through camera-equipped mobile devices.

The study proposes three CNN architectures: two hybrid models (CNN+SVM and CNN+kNN) and a ResNet-50 network. Various preprocessing steps using detection techniques including Gaussian filtering were applied to a dataset of 7,000 pill images from the National Library of Medicine.

The proposed CNN+kNN architecture achieved 90.8% accuracy, a roughly 10% improvement over existing methods, with a runtime of approximately 1 millisecond per execution.

I co-authored this paper alongside faculty at Rochester Institute of Technology, Dubai. I designed the experimental methodology, selected and tuned the CNN architectures, ran model training across the three approaches, and led the evaluation against the benchmark dataset.

The core engineering challenge was pushing accuracy past the 80% ceiling of existing methods while keeping inference fast enough for mobile use. The CNN+kNN hybrid hit 90.8% at ~1ms per execution — the architecture choice mattered as much as the training.