MobilePill: Accurate Pill Image Classification via Deep Learning on Mobile

It is cumbersome for older people to search for prescription drugs via the Internet. Thus, recent studies use mobile to help this inconvenience. One of them is to provide the details of pill information prescribed through mobile. However, previous studies have shown many difficulties in providing drug information through mobile. In particular, it is more challenging to provide the pill information after classifying drug image in terms of accuracy. In this paper, we tackle the problem of accuracy, whether a given pill image is correctly classified via mobile phone. In this circumstance, we develop an approach that uses a hybrid CNN model to classify various pill images. This model adopts approaches to classify different properties in pill images with the same shape. In reality, we designed a system called "MobilePill," which is a hybrid model with each different independent model, and have each different training sets according to dosage forms. In extensive experiments, "MobilePill" achieved 73.39% accuracy for the case of one-sided, which was 39.25% higher than previous models. This result provides the possibilities that can be applied in real-world.